├── DAVIS ├── Annotations │ └── 480p │ │ └── car-shadow │ │ ├── 00000.png │ │ ├── 00001.png │ │ ├── 00002.png │ │ ├── 00003.png │ │ ├── 00004.png │ │ ├── 00005.png │ │ ├── 00006.png │ │ ├── 00007.png │ │ ├── 00008.png │ │ ├── 00009.png │ │ ├── 00010.png │ │ ├── 00011.png │ │ ├── 00012.png │ │ ├── 00013.png │ │ ├── 00014.png │ │ ├── 00015.png │ │ ├── 00016.png │ │ ├── 00017.png │ │ ├── 00018.png │ │ ├── 00019.png │ │ ├── 00020.png │ │ ├── 00021.png │ │ ├── 00022.png │ │ ├── 00023.png │ │ ├── 00024.png │ │ ├── 00025.png │ │ ├── 00026.png │ │ ├── 00027.png │ │ ├── 00028.png │ │ ├── 00029.png │ │ ├── 00030.png │ │ ├── 00031.png │ │ ├── 00032.png │ │ ├── 00033.png │ │ ├── 00034.png │ │ ├── 00035.png │ │ ├── 00036.png │ │ ├── 00037.png │ │ ├── 00038.png │ │ └── 00039.png └── JPEGImages │ └── 480p │ └── car-shadow │ ├── 00000.jpg │ ├── 00001.jpg │ ├── 00002.jpg │ ├── 00003.jpg │ ├── 00004.jpg │ ├── 00005.jpg │ ├── 00006.jpg │ ├── 00007.jpg │ ├── 00008.jpg │ ├── 00009.jpg │ ├── 00010.jpg │ ├── 00011.jpg │ ├── 00012.jpg │ ├── 00013.jpg │ ├── 00014.jpg │ ├── 00015.jpg │ ├── 00016.jpg │ ├── 00017.jpg │ ├── 00018.jpg │ ├── 00019.jpg │ ├── 00020.jpg │ ├── 00021.jpg │ ├── 00022.jpg │ ├── 00023.jpg │ ├── 00024.jpg │ ├── 00025.jpg │ ├── 00026.jpg │ ├── 00027.jpg │ ├── 00028.jpg │ ├── 00029.jpg │ ├── 00030.jpg │ ├── 00031.jpg │ ├── 00032.jpg │ ├── 00033.jpg │ ├── 00034.jpg │ ├── 00035.jpg │ ├── 00036.jpg │ ├── 00037.jpg │ ├── 00038.jpg │ └── 00039.jpg ├── LICENSE ├── README.md ├── dataset.py ├── doc └── ims │ └── osvos.png ├── models └── .gitignore ├── osvos.py ├── osvos_demo.py ├── osvos_parent_demo.py ├── requirements.txt └── train_parent.txt /DAVIS/Annotations/480p/car-shadow/00000.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00000.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00001.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00001.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00002.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00002.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00003.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00003.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00004.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00004.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00005.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00005.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00006.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00006.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00007.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00007.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00008.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00008.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00009.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00009.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00010.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00010.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00011.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00011.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00012.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00012.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00013.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00013.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00014.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00014.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00015.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00015.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00016.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00016.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00017.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00017.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00018.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00018.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00019.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00019.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00020.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00020.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00021.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00021.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00022.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00022.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00023.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00023.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00024.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00024.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00025.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00025.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00026.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00026.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00027.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00027.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00028.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00028.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00029.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00029.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00030.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00030.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00031.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00031.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00032.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00032.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00033.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00033.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00034.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00034.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00035.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00035.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00036.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00036.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00037.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00037.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00038.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00038.png -------------------------------------------------------------------------------- /DAVIS/Annotations/480p/car-shadow/00039.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/Annotations/480p/car-shadow/00039.png -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00000.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00000.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00001.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00001.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00002.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00002.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00003.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00003.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00004.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00004.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00005.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00005.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00006.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00006.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00007.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00007.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00008.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00008.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00009.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00009.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00010.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00010.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00011.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00011.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00012.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00012.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00013.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00013.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00014.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00014.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00015.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00015.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00016.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00016.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00017.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00017.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00018.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00018.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00019.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00019.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00020.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00020.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00021.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00021.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00022.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00022.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00023.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00023.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00024.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00024.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00025.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00025.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00026.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00026.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00027.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00027.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00028.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00028.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00029.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00029.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00030.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00030.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00031.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00031.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00032.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00032.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00033.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00033.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00034.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00034.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00035.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00035.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00036.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00036.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00037.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00037.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00038.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00038.jpg -------------------------------------------------------------------------------- /DAVIS/JPEGImages/480p/car-shadow/00039.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/DAVIS/JPEGImages/480p/car-shadow/00039.jpg -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # OSVOS: One-Shot Video Object Segmentation 2 | Check our [project page](http://www.vision.ee.ethz.ch/~cvlsegmentation/osvos) for additional information. 3 | ![OSVOS](doc/ims/osvos.png) 4 | 5 | OSVOS is a method that tackles the task of semi-supervised video object segmentation. It is based on a fully-convolutional neural network architecture that is able to successively transfer generic semantic information, learned on ImageNet, to the task of foreground segmentation, and finally to learning the appearance of a single annotated object of the test sequence (hence one-shot). Experiments on DAVIS 2016 show that OSVOS is faster than currently available techniques and improves the state of the art by a significant margin (79.8% vs 68.0%). 6 | 7 | 8 | This TensorFlow code is a posteriori implementation of OSVOS and it does not contain the boundary snapping branch. The results published in the paper were obtained using the Caffe version that can be found at [OSVOS-caffe](https://github.com/kmaninis/OSVOS-caffe). 9 | 10 | #### NEW: PyTorch implementation also available: [OSVOS-PyTorch](https://github.com/kmaninis/OSVOS-PyTorch)! 11 | 12 | ### Installation: 13 | 1. Clone the OSVOS-TensorFlow repository 14 | ```Shell 15 | git clone https://github.com/scaelles/OSVOS-TensorFlow.git 16 | ``` 17 | 2. Install if necessary the required dependencies: 18 | 19 | - Python 2.7, Python 3 (thanks to [@xoltar](https://github.com/xoltar)) 20 | - Tensorflow r1.0 or higher (`pip install tensorflow-gpu`) along with standard [dependencies](https://www.tensorflow.org/install/install_linux) 21 | - Other python dependencies: PIL (Pillow version), numpy, scipy, matplotlib, six 22 | 23 | 3. Download the parent model from [here](https://data.vision.ee.ethz.ch/csergi/share/OSVOS/OSVOS_parent_model.zip) (55 MB) and unzip it under `models/` (It should create a folder named 'OSVOS_parent'). 24 | 25 | 4. All the steps to re-train OSVOS are provided in this repository. In case you would like to test with the pre-trained models, you can download them from [here](https://data.vision.ee.ethz.ch/csergi/share/OSVOS/OSVOS_pre-trained_models.zip) (2.2GB) and unzip them under `models/` (It should create a folder for every model). 26 | 27 | ### Demo online training and testing 28 | 1. Edit in file `osvos_demo.py` the 'User defined parameters' (eg. gpu_id, train_model, etc). 29 | 30 | 2. Run `python osvos_demo.py`. 31 | 32 | It is possible to work with all sequences of DAVIS 2016 just by creating a soft link (`ln -s /path/to/DAVIS/ DAVIS`) in the root folder of the project. 33 | 34 | ### Training the parent network (optional) 35 | 1. All the training sequences of DAVIS 2016 are required to train the parent model, thus download it from [here](https://graphics.ethz.ch/Downloads/Data/Davis/DAVIS-data.zip) if you don't have it. 36 | 2. Place the dataset in this repository or create a soft link to it (`ln -s /path/to/DAVIS/ DAVIS`) if you have it somewhere else. 37 | 3. Download the VGG 16 model trained on Imagenet from the TF model zoo from [here](http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz). 38 | 4. Place the vgg_16.ckpt file inside `models/`. 39 | 5. Edit the 'User defined parameters' (eg. gpu_id) in file `osvos_parent_demo.py`. 40 | 6. Run `python osvos_parent_demo.py`. This step takes 20 hours to train (Titan-X Pascal), and ~15GB for loading data and online data augmentation. Change dataset.py accordingly, to adjust to a less memory-intensive setup. 41 | 42 | Have a happy training! 43 | 44 | ### Citation: 45 | @Inproceedings{Cae+17, 46 | Title = {One-Shot Video Object Segmentation}, 47 | Author = {S. Caelles and K.K. Maninis and J. Pont-Tuset and L. Leal-Taix\'e and D. Cremers and L. {Van Gool}}, 48 | Booktitle = {Computer Vision and Pattern Recognition (CVPR)}, 49 | Year = {2017} 50 | } 51 | If you encounter any problems with the code, want to report bugs, etc. please contact me at scaelles[at]vision[dot]ee[dot]ethz[dot]ch. 52 | -------------------------------------------------------------------------------- /dataset.py: -------------------------------------------------------------------------------- 1 | """ 2 | Sergi Caelles (scaelles@vision.ee.ethz.ch) 3 | 4 | This file is part of the OSVOS paper presented in: 5 | Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, Luc Van Gool 6 | One-Shot Video Object Segmentation 7 | CVPR 2017 8 | Please consider citing the paper if you use this code. 9 | """ 10 | from PIL import Image 11 | import os 12 | import numpy as np 13 | import sys 14 | 15 | 16 | class Dataset: 17 | def __init__(self, train_list, test_list, database_root, store_memory=True, data_aug=False): 18 | """Initialize the Dataset object 19 | Args: 20 | train_list: TXT file or list with the paths of the images to use for training (Images must be between 0 and 255) 21 | test_list: TXT file or list with the paths of the images to use for testing (Images must be between 0 and 255) 22 | database_root: Path to the root of the Database 23 | store_memory: True stores all the training images, False loads at runtime the images 24 | Returns: 25 | """ 26 | if not store_memory and data_aug: 27 | sys.stderr.write('Online data augmentation not supported when the data is not stored in memory!') 28 | sys.exit() 29 | # Define types of data augmentation 30 | data_aug_scales = [0.5, 0.8, 1] 31 | data_aug_flip = True 32 | 33 | # Load training images (path) and labels 34 | print('Started loading files...') 35 | if not isinstance(train_list, list) and train_list is not None: 36 | with open(train_list) as t: 37 | train_paths = t.readlines() 38 | elif isinstance(train_list, list): 39 | train_paths = train_list 40 | else: 41 | train_paths = [] 42 | if not isinstance(test_list, list) and test_list is not None: 43 | with open(test_list) as t: 44 | test_paths = t.readlines() 45 | elif isinstance(test_list, list): 46 | test_paths = test_list 47 | else: 48 | test_paths = [] 49 | self.images_train = [] 50 | self.images_train_path = [] 51 | self.labels_train = [] 52 | self.labels_train_path = [] 53 | for idx, line in enumerate(train_paths): 54 | if store_memory: 55 | img = Image.open(os.path.join(database_root, str(line.split()[0]))) 56 | img.load() 57 | label = Image.open(os.path.join(database_root, str(line.split()[1]))) 58 | label.load() 59 | label = label.split()[0] 60 | if data_aug: 61 | if idx == 0: sys.stdout.write('Performing the data augmentation') 62 | for scale in data_aug_scales: 63 | img_size = tuple([int(img.size[0] * scale), int(img.size[1] * scale)]) 64 | img_sc = img.resize(img_size) 65 | label_sc = label.resize(img_size) 66 | self.images_train.append(np.array(img_sc, dtype=np.uint8)) 67 | self.labels_train.append(np.array(label_sc, dtype=np.uint8)) 68 | if data_aug_flip: 69 | img_sc_fl = img_sc.transpose(Image.FLIP_LEFT_RIGHT) 70 | label_sc_fl = label_sc.transpose(Image.FLIP_LEFT_RIGHT) 71 | self.images_train.append(np.array(img_sc_fl, dtype=np.uint8)) 72 | self.labels_train.append(np.array(label_sc_fl, dtype=np.uint8)) 73 | else: 74 | if idx == 0: sys.stdout.write('Loading the data') 75 | self.images_train.append(np.array(img, dtype=np.uint8)) 76 | self.labels_train.append(np.array(label, dtype=np.uint8)) 77 | if (idx + 1) % 50 == 0: 78 | sys.stdout.write('.') 79 | self.images_train_path.append(os.path.join(database_root, str(line.split()[0]))) 80 | self.labels_train_path.append(os.path.join(database_root, str(line.split()[1]))) 81 | sys.stdout.write('\n') 82 | self.images_train_path = np.array(self.images_train_path) 83 | self.labels_train_path = np.array(self.labels_train_path) 84 | 85 | # Load testing images (path) and labels 86 | self.images_test = [] 87 | self.images_test_path = [] 88 | for idx, line in enumerate(test_paths): 89 | if store_memory: 90 | self.images_test.append(np.array(Image.open(os.path.join(database_root, str(line.split()[0]))), 91 | dtype=np.uint8)) 92 | if (idx + 1) % 1000 == 0: 93 | print('Loaded ' + str(idx) + ' test images') 94 | self.images_test_path.append(os.path.join(database_root, str(line.split()[0]))) 95 | print('Done initializing Dataset') 96 | 97 | # Init parameters 98 | self.train_ptr = 0 99 | self.test_ptr = 0 100 | self.train_size = max(len(self.images_train_path), len(self.images_train)) 101 | self.test_size = len(self.images_test_path) 102 | self.train_idx = np.arange(self.train_size) 103 | np.random.shuffle(self.train_idx) 104 | self.store_memory = store_memory 105 | 106 | def next_batch(self, batch_size, phase): 107 | """Get next batch of image (path) and labels 108 | Args: 109 | batch_size: Size of the batch 110 | phase: Possible options:'train' or 'test' 111 | Returns in training: 112 | images: List of images paths if store_memory=False, List of Numpy arrays of the images if store_memory=True 113 | labels: List of labels paths if store_memory=False, List of Numpy arrays of the labels if store_memory=True 114 | Returns in testing: 115 | images: None if store_memory=False, Numpy array of the image if store_memory=True 116 | path: List of image paths 117 | """ 118 | if phase == 'train': 119 | if self.train_ptr + batch_size < self.train_size: 120 | idx = np.array(self.train_idx[self.train_ptr:self.train_ptr + batch_size]) 121 | if self.store_memory: 122 | images = [self.images_train[l] for l in idx] 123 | labels = [self.labels_train[l] for l in idx] 124 | else: 125 | images = [self.images_train_path[l] for l in idx] 126 | labels = [self.labels_train_path[l] for l in idx] 127 | self.train_ptr += batch_size 128 | else: 129 | old_idx = np.array(self.train_idx[self.train_ptr:]) 130 | np.random.shuffle(self.train_idx) 131 | new_ptr = (self.train_ptr + batch_size) % self.train_size 132 | idx = np.array(self.train_idx[:new_ptr]) 133 | if self.store_memory: 134 | images_1 = [self.images_train[l] for l in old_idx] 135 | labels_1 = [self.labels_train[l] for l in old_idx] 136 | images_2 = [self.images_train[l] for l in idx] 137 | labels_2 = [self.labels_train[l] for l in idx] 138 | else: 139 | images_1 = [self.images_train_path[l] for l in old_idx] 140 | labels_1 = [self.labels_train_path[l] for l in old_idx] 141 | images_2 = [self.images_train_path[l] for l in idx] 142 | labels_2 = [self.labels_train_path[l] for l in idx] 143 | images = images_1 + images_2 144 | labels = labels_1 + labels_2 145 | self.train_ptr = new_ptr 146 | return images, labels 147 | elif phase == 'test': 148 | images = None 149 | if self.test_ptr + batch_size < self.test_size: 150 | if self.store_memory: 151 | images = self.images_test[self.test_ptr:self.test_ptr + batch_size] 152 | paths = self.images_test_path[self.test_ptr:self.test_ptr + batch_size] 153 | self.test_ptr += batch_size 154 | else: 155 | new_ptr = (self.test_ptr + batch_size) % self.test_size 156 | if self.store_memory: 157 | images = self.images_test[self.test_ptr:] + self.images_test[:new_ptr] 158 | paths = self.images_test_path[self.test_ptr:] + self.images_test_path[:new_ptr] 159 | self.test_ptr = new_ptr 160 | return images, paths 161 | else: 162 | return None, None 163 | 164 | def get_train_size(self): 165 | return self.train_size 166 | 167 | def get_test_size(self): 168 | return self.test_size 169 | 170 | def train_img_size(self): 171 | width, height = Image.open(self.images_train[self.train_ptr]).size 172 | return height, width 173 | -------------------------------------------------------------------------------- /doc/ims/osvos.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/scaelles/OSVOS-TensorFlow/6ef61d1f523296b95974a6dafac01e94bf2fde7d/doc/ims/osvos.png -------------------------------------------------------------------------------- /models/.gitignore: -------------------------------------------------------------------------------- 1 | # Ignore everything in this directory 2 | * 3 | # Except this file 4 | !.gitignore 5 | -------------------------------------------------------------------------------- /osvos.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | """ 3 | Sergi Caelles (scaelles@vision.ee.ethz.ch) 4 | 5 | This file is part of the OSVOS paper presented in: 6 | Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, Luc Van Gool 7 | One-Shot Video Object Segmentation 8 | CVPR 2017 9 | Please consider citing the paper if you use this code. 10 | """ 11 | import tensorflow as tf 12 | import numpy as np 13 | from tensorflow.contrib.layers.python.layers import utils 14 | import sys 15 | from datetime import datetime 16 | import os 17 | import scipy.misc 18 | from PIL import Image 19 | import six 20 | 21 | slim = tf.contrib.slim 22 | 23 | 24 | def osvos_arg_scope(weight_decay=0.0002): 25 | """Defines the OSVOS arg scope. 26 | Args: 27 | weight_decay: The l2 regularization coefficient. 28 | Returns: 29 | An arg_scope. 30 | """ 31 | with slim.arg_scope([slim.conv2d, slim.convolution2d_transpose], 32 | activation_fn=tf.nn.relu, 33 | weights_initializer=tf.random_normal_initializer(stddev=0.001), 34 | weights_regularizer=slim.l2_regularizer(weight_decay), 35 | biases_initializer=tf.zeros_initializer(), 36 | biases_regularizer=None, 37 | padding='SAME') as arg_sc: 38 | return arg_sc 39 | 40 | 41 | def crop_features(feature, out_size): 42 | """Crop the center of a feature map 43 | Args: 44 | feature: Feature map to crop 45 | out_size: Size of the output feature map 46 | Returns: 47 | Tensor that performs the cropping 48 | """ 49 | up_size = tf.shape(feature) 50 | ini_w = tf.div(tf.subtract(up_size[1], out_size[1]), 2) 51 | ini_h = tf.div(tf.subtract(up_size[2], out_size[2]), 2) 52 | slice_input = tf.slice(feature, (0, ini_w, ini_h, 0), (-1, out_size[1], out_size[2], -1)) 53 | # slice_input = tf.slice(feature, (0, ini_w, ini_w, 0), (-1, out_size[1], out_size[2], -1)) # Caffe cropping way 54 | return tf.reshape(slice_input, [int(feature.get_shape()[0]), out_size[1], out_size[2], int(feature.get_shape()[3])]) 55 | 56 | 57 | def osvos(inputs, scope='osvos'): 58 | """Defines the OSVOS network 59 | Args: 60 | inputs: Tensorflow placeholder that contains the input image 61 | scope: Scope name for the network 62 | Returns: 63 | net: Output Tensor of the network 64 | end_points: Dictionary with all Tensors of the network 65 | """ 66 | im_size = tf.shape(inputs) 67 | 68 | with tf.variable_scope(scope, 'osvos', [inputs]) as sc: 69 | end_points_collection = sc.name + '_end_points' 70 | # Collect outputs of all intermediate layers. 71 | with slim.arg_scope([slim.conv2d, slim.max_pool2d], 72 | padding='SAME', 73 | outputs_collections=end_points_collection): 74 | net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1') 75 | net = slim.max_pool2d(net, [2, 2], scope='pool1') 76 | net_2 = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2') 77 | net = slim.max_pool2d(net_2, [2, 2], scope='pool2') 78 | net_3 = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3') 79 | net = slim.max_pool2d(net_3, [2, 2], scope='pool3') 80 | net_4 = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4') 81 | net = slim.max_pool2d(net_4, [2, 2], scope='pool4') 82 | net_5 = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5') 83 | 84 | # Get side outputs of the network 85 | with slim.arg_scope([slim.conv2d], 86 | activation_fn=None): 87 | side_2 = slim.conv2d(net_2, 16, [3, 3], scope='conv2_2_16') 88 | side_3 = slim.conv2d(net_3, 16, [3, 3], scope='conv3_3_16') 89 | side_4 = slim.conv2d(net_4, 16, [3, 3], scope='conv4_3_16') 90 | side_5 = slim.conv2d(net_5, 16, [3, 3], scope='conv5_3_16') 91 | 92 | # Supervise side outputs 93 | side_2_s = slim.conv2d(side_2, 1, [1, 1], scope='score-dsn_2') 94 | side_3_s = slim.conv2d(side_3, 1, [1, 1], scope='score-dsn_3') 95 | side_4_s = slim.conv2d(side_4, 1, [1, 1], scope='score-dsn_4') 96 | side_5_s = slim.conv2d(side_5, 1, [1, 1], scope='score-dsn_5') 97 | with slim.arg_scope([slim.convolution2d_transpose], 98 | activation_fn=None, biases_initializer=None, padding='VALID', 99 | outputs_collections=end_points_collection, trainable=False): 100 | # Side outputs 101 | side_2_s = slim.convolution2d_transpose(side_2_s, 1, 4, 2, scope='score-dsn_2-up') 102 | side_2_s = crop_features(side_2_s, im_size) 103 | utils.collect_named_outputs(end_points_collection, 'osvos/score-dsn_2-cr', side_2_s) 104 | side_3_s = slim.convolution2d_transpose(side_3_s, 1, 8, 4, scope='score-dsn_3-up') 105 | side_3_s = crop_features(side_3_s, im_size) 106 | utils.collect_named_outputs(end_points_collection, 'osvos/score-dsn_3-cr', side_3_s) 107 | side_4_s = slim.convolution2d_transpose(side_4_s, 1, 16, 8, scope='score-dsn_4-up') 108 | side_4_s = crop_features(side_4_s, im_size) 109 | utils.collect_named_outputs(end_points_collection, 'osvos/score-dsn_4-cr', side_4_s) 110 | side_5_s = slim.convolution2d_transpose(side_5_s, 1, 32, 16, scope='score-dsn_5-up') 111 | side_5_s = crop_features(side_5_s, im_size) 112 | utils.collect_named_outputs(end_points_collection, 'osvos/score-dsn_5-cr', side_5_s) 113 | 114 | # Main output 115 | side_2_f = slim.convolution2d_transpose(side_2, 16, 4, 2, scope='score-multi2-up') 116 | side_2_f = crop_features(side_2_f, im_size) 117 | utils.collect_named_outputs(end_points_collection, 'osvos/side-multi2-cr', side_2_f) 118 | side_3_f = slim.convolution2d_transpose(side_3, 16, 8, 4, scope='score-multi3-up') 119 | side_3_f = crop_features(side_3_f, im_size) 120 | utils.collect_named_outputs(end_points_collection, 'osvos/side-multi3-cr', side_3_f) 121 | side_4_f = slim.convolution2d_transpose(side_4, 16, 16, 8, scope='score-multi4-up') 122 | side_4_f = crop_features(side_4_f, im_size) 123 | utils.collect_named_outputs(end_points_collection, 'osvos/side-multi4-cr', side_4_f) 124 | side_5_f = slim.convolution2d_transpose(side_5, 16, 32, 16, scope='score-multi5-up') 125 | side_5_f = crop_features(side_5_f, im_size) 126 | utils.collect_named_outputs(end_points_collection, 'osvos/side-multi5-cr', side_5_f) 127 | concat_side = tf.concat([side_2_f, side_3_f, side_4_f, side_5_f], axis=3) 128 | 129 | net = slim.conv2d(concat_side, 1, [1, 1], scope='upscore-fuse') 130 | 131 | end_points = slim.utils.convert_collection_to_dict(end_points_collection) 132 | return net, end_points 133 | 134 | 135 | def upsample_filt(size): 136 | factor = (size + 1) // 2 137 | if size % 2 == 1: 138 | center = factor - 1 139 | else: 140 | center = factor - 0.5 141 | og = np.ogrid[:size, :size] 142 | return (1 - abs(og[0] - center) / factor) * \ 143 | (1 - abs(og[1] - center) / factor) 144 | 145 | 146 | # Set deconvolutional layers to compute bilinear interpolation 147 | def interp_surgery(variables): 148 | interp_tensors = [] 149 | for v in variables: 150 | if '-up' in v.name: 151 | h, w, k, m = v.get_shape() 152 | tmp = np.zeros((m, k, h, w)) 153 | if m != k: 154 | raise ValueError('input + output channels need to be the same') 155 | if h != w: 156 | raise ValueError('filters need to be square') 157 | up_filter = upsample_filt(int(h)) 158 | tmp[range(m), range(k), :, :] = up_filter 159 | interp_tensors.append(tf.assign(v, tmp.transpose((2, 3, 1, 0)), validate_shape=True, use_locking=True)) 160 | return interp_tensors 161 | 162 | 163 | # TO DO: Move preprocessing into Tensorflow 164 | def preprocess_img(image): 165 | """Preprocess the image to adapt it to network requirements 166 | Args: 167 | Image we want to input the network (W,H,3) numpy array 168 | Returns: 169 | Image ready to input the network (1,W,H,3) 170 | """ 171 | if type(image) is not np.ndarray: 172 | image = np.array(Image.open(image), dtype=np.uint8) 173 | in_ = image[:, :, ::-1] 174 | in_ = np.subtract(in_, np.array((104.00699, 116.66877, 122.67892), dtype=np.float32)) 175 | # in_ = tf.subtract(tf.cast(in_, tf.float32), np.array((104.00699, 116.66877, 122.67892), dtype=np.float32)) 176 | in_ = np.expand_dims(in_, axis=0) 177 | # in_ = tf.expand_dims(in_, 0) 178 | return in_ 179 | 180 | 181 | # TO DO: Move preprocessing into Tensorflow 182 | def preprocess_labels(label): 183 | """Preprocess the labels to adapt them to the loss computation requirements 184 | Args: 185 | Label corresponding to the input image (W,H) numpy array 186 | Returns: 187 | Label ready to compute the loss (1,W,H,1) 188 | """ 189 | if type(label) is not np.ndarray: 190 | label = np.array(Image.open(label).split()[0], dtype=np.uint8) 191 | max_mask = np.max(label) * 0.5 192 | label = np.greater(label, max_mask) 193 | label = np.expand_dims(np.expand_dims(label, axis=0), axis=3) 194 | # label = tf.cast(np.array(label), tf.float32) 195 | # max_mask = tf.multiply(tf.reduce_max(label), 0.5) 196 | # label = tf.cast(tf.greater(label, max_mask), tf.float32) 197 | # label = tf.expand_dims(tf.expand_dims(label, 0), 3) 198 | return label 199 | 200 | 201 | def load_vgg_imagenet(ckpt_path): 202 | """Initialize the network parameters from the VGG-16 pre-trained model provided by TF-SLIM 203 | Args: 204 | Path to the checkpoint 205 | Returns: 206 | Function that takes a session and initializes the network 207 | """ 208 | reader = tf.train.NewCheckpointReader(ckpt_path) 209 | var_to_shape_map = reader.get_variable_to_shape_map() 210 | vars_corresp = dict() 211 | for v in var_to_shape_map: 212 | if "conv" in v: 213 | vars_corresp[v] = slim.get_model_variables(v.replace("vgg_16", "osvos"))[0] 214 | init_fn = slim.assign_from_checkpoint_fn( 215 | ckpt_path, 216 | vars_corresp) 217 | return init_fn 218 | 219 | 220 | def class_balanced_cross_entropy_loss(output, label): 221 | """Define the class balanced cross entropy loss to train the network 222 | Args: 223 | output: Output of the network 224 | label: Ground truth label 225 | Returns: 226 | Tensor that evaluates the loss 227 | """ 228 | 229 | labels = tf.cast(tf.greater(label, 0.5), tf.float32) 230 | 231 | num_labels_pos = tf.reduce_sum(labels) 232 | num_labels_neg = tf.reduce_sum(1.0 - labels) 233 | num_total = num_labels_pos + num_labels_neg 234 | 235 | output_gt_zero = tf.cast(tf.greater_equal(output, 0), tf.float32) 236 | loss_val = tf.multiply(output, (labels - output_gt_zero)) - tf.log( 237 | 1 + tf.exp(output - 2 * tf.multiply(output, output_gt_zero))) 238 | 239 | loss_pos = tf.reduce_sum(-tf.multiply(labels, loss_val)) 240 | loss_neg = tf.reduce_sum(-tf.multiply(1.0 - labels, loss_val)) 241 | 242 | final_loss = num_labels_neg / num_total * loss_pos + num_labels_pos / num_total * loss_neg 243 | 244 | return final_loss 245 | 246 | 247 | def class_balanced_cross_entropy_loss_theoretical(output, label): 248 | """Theoretical version of the class balanced cross entropy loss to train the network (Produces unstable results) 249 | Args: 250 | output: Output of the network 251 | label: Ground truth label 252 | Returns: 253 | Tensor that evaluates the loss 254 | """ 255 | output = tf.nn.sigmoid(output) 256 | 257 | labels_pos = tf.cast(tf.greater(label, 0), tf.float32) 258 | labels_neg = tf.cast(tf.less(label, 1), tf.float32) 259 | 260 | num_labels_pos = tf.reduce_sum(labels_pos) 261 | num_labels_neg = tf.reduce_sum(labels_neg) 262 | num_total = num_labels_pos + num_labels_neg 263 | 264 | loss_pos = tf.reduce_sum(tf.multiply(labels_pos, tf.log(output + 0.00001))) 265 | loss_neg = tf.reduce_sum(tf.multiply(labels_neg, tf.log(1 - output + 0.00001))) 266 | 267 | final_loss = -num_labels_neg / num_total * loss_pos - num_labels_pos / num_total * loss_neg 268 | 269 | return final_loss 270 | 271 | 272 | def load_caffe_weights(weights_path): 273 | """Initialize the network parameters from a .npy caffe weights file 274 | Args: 275 | Path to the .npy file containing the value of the network parameters 276 | Returns: 277 | Function that takes a session and initializes the network 278 | """ 279 | osvos_weights = np.load(weights_path).item() 280 | vars_corresp = dict() 281 | vars_corresp['osvos/conv1/conv1_1/weights'] = osvos_weights['conv1_1_w'] 282 | vars_corresp['osvos/conv1/conv1_1/biases'] = osvos_weights['conv1_1_b'] 283 | vars_corresp['osvos/conv1/conv1_2/weights'] = osvos_weights['conv1_2_w'] 284 | vars_corresp['osvos/conv1/conv1_2/biases'] = osvos_weights['conv1_2_b'] 285 | 286 | vars_corresp['osvos/conv2/conv2_1/weights'] = osvos_weights['conv2_1_w'] 287 | vars_corresp['osvos/conv2/conv2_1/biases'] = osvos_weights['conv2_1_b'] 288 | vars_corresp['osvos/conv2/conv2_2/weights'] = osvos_weights['conv2_2_w'] 289 | vars_corresp['osvos/conv2/conv2_2/biases'] = osvos_weights['conv2_2_b'] 290 | 291 | vars_corresp['osvos/conv3/conv3_1/weights'] = osvos_weights['conv3_1_w'] 292 | vars_corresp['osvos/conv3/conv3_1/biases'] = osvos_weights['conv3_1_b'] 293 | vars_corresp['osvos/conv3/conv3_2/weights'] = osvos_weights['conv3_2_w'] 294 | vars_corresp['osvos/conv3/conv3_2/biases'] = osvos_weights['conv3_2_b'] 295 | vars_corresp['osvos/conv3/conv3_3/weights'] = osvos_weights['conv3_3_w'] 296 | vars_corresp['osvos/conv3/conv3_3/biases'] = osvos_weights['conv3_3_b'] 297 | 298 | vars_corresp['osvos/conv4/conv4_1/weights'] = osvos_weights['conv4_1_w'] 299 | vars_corresp['osvos/conv4/conv4_1/biases'] = osvos_weights['conv4_1_b'] 300 | vars_corresp['osvos/conv4/conv4_2/weights'] = osvos_weights['conv4_2_w'] 301 | vars_corresp['osvos/conv4/conv4_2/biases'] = osvos_weights['conv4_2_b'] 302 | vars_corresp['osvos/conv4/conv4_3/weights'] = osvos_weights['conv4_3_w'] 303 | vars_corresp['osvos/conv4/conv4_3/biases'] = osvos_weights['conv4_3_b'] 304 | 305 | vars_corresp['osvos/conv5/conv5_1/weights'] = osvos_weights['conv5_1_w'] 306 | vars_corresp['osvos/conv5/conv5_1/biases'] = osvos_weights['conv5_1_b'] 307 | vars_corresp['osvos/conv5/conv5_2/weights'] = osvos_weights['conv5_2_w'] 308 | vars_corresp['osvos/conv5/conv5_2/biases'] = osvos_weights['conv5_2_b'] 309 | vars_corresp['osvos/conv5/conv5_3/weights'] = osvos_weights['conv5_3_w'] 310 | vars_corresp['osvos/conv5/conv5_3/biases'] = osvos_weights['conv5_3_b'] 311 | 312 | vars_corresp['osvos/conv2_2_16/weights'] = osvos_weights['conv2_2_16_w'] 313 | vars_corresp['osvos/conv2_2_16/biases'] = osvos_weights['conv2_2_16_b'] 314 | vars_corresp['osvos/conv3_3_16/weights'] = osvos_weights['conv3_3_16_w'] 315 | vars_corresp['osvos/conv3_3_16/biases'] = osvos_weights['conv3_3_16_b'] 316 | vars_corresp['osvos/conv4_3_16/weights'] = osvos_weights['conv4_3_16_w'] 317 | vars_corresp['osvos/conv4_3_16/biases'] = osvos_weights['conv4_3_16_b'] 318 | vars_corresp['osvos/conv5_3_16/weights'] = osvos_weights['conv5_3_16_w'] 319 | vars_corresp['osvos/conv5_3_16/biases'] = osvos_weights['conv5_3_16_b'] 320 | 321 | vars_corresp['osvos/score-dsn_2/weights'] = osvos_weights['score-dsn_2_w'] 322 | vars_corresp['osvos/score-dsn_2/biases'] = osvos_weights['score-dsn_2_b'] 323 | vars_corresp['osvos/score-dsn_3/weights'] = osvos_weights['score-dsn_3_w'] 324 | vars_corresp['osvos/score-dsn_3/biases'] = osvos_weights['score-dsn_3_b'] 325 | vars_corresp['osvos/score-dsn_4/weights'] = osvos_weights['score-dsn_4_w'] 326 | vars_corresp['osvos/score-dsn_4/biases'] = osvos_weights['score-dsn_4_b'] 327 | vars_corresp['osvos/score-dsn_5/weights'] = osvos_weights['score-dsn_5_w'] 328 | vars_corresp['osvos/score-dsn_5/biases'] = osvos_weights['score-dsn_5_b'] 329 | 330 | vars_corresp['osvos/upscore-fuse/weights'] = osvos_weights['new-score-weighting_w'] 331 | vars_corresp['osvos/upscore-fuse/biases'] = osvos_weights['new-score-weighting_b'] 332 | return slim.assign_from_values_fn(vars_corresp) 333 | 334 | 335 | def parameter_lr(): 336 | """Specify the relative learning rate for every parameter. The final learning rate 337 | in every parameter will be the one defined here multiplied by the global one 338 | Args: 339 | Returns: 340 | Dictionary with the relative learning rate for every parameter 341 | """ 342 | 343 | vars_corresp = dict() 344 | vars_corresp['osvos/conv1/conv1_1/weights'] = 1 345 | vars_corresp['osvos/conv1/conv1_1/biases'] = 2 346 | vars_corresp['osvos/conv1/conv1_2/weights'] = 1 347 | vars_corresp['osvos/conv1/conv1_2/biases'] = 2 348 | 349 | vars_corresp['osvos/conv2/conv2_1/weights'] = 1 350 | vars_corresp['osvos/conv2/conv2_1/biases'] = 2 351 | vars_corresp['osvos/conv2/conv2_2/weights'] = 1 352 | vars_corresp['osvos/conv2/conv2_2/biases'] = 2 353 | 354 | vars_corresp['osvos/conv3/conv3_1/weights'] = 1 355 | vars_corresp['osvos/conv3/conv3_1/biases'] = 2 356 | vars_corresp['osvos/conv3/conv3_2/weights'] = 1 357 | vars_corresp['osvos/conv3/conv3_2/biases'] = 2 358 | vars_corresp['osvos/conv3/conv3_3/weights'] = 1 359 | vars_corresp['osvos/conv3/conv3_3/biases'] = 2 360 | 361 | vars_corresp['osvos/conv4/conv4_1/weights'] = 1 362 | vars_corresp['osvos/conv4/conv4_1/biases'] = 2 363 | vars_corresp['osvos/conv4/conv4_2/weights'] = 1 364 | vars_corresp['osvos/conv4/conv4_2/biases'] = 2 365 | vars_corresp['osvos/conv4/conv4_3/weights'] = 1 366 | vars_corresp['osvos/conv4/conv4_3/biases'] = 2 367 | 368 | vars_corresp['osvos/conv5/conv5_1/weights'] = 1 369 | vars_corresp['osvos/conv5/conv5_1/biases'] = 2 370 | vars_corresp['osvos/conv5/conv5_2/weights'] = 1 371 | vars_corresp['osvos/conv5/conv5_2/biases'] = 2 372 | vars_corresp['osvos/conv5/conv5_3/weights'] = 1 373 | vars_corresp['osvos/conv5/conv5_3/biases'] = 2 374 | 375 | vars_corresp['osvos/conv2_2_16/weights'] = 1 376 | vars_corresp['osvos/conv2_2_16/biases'] = 2 377 | vars_corresp['osvos/conv3_3_16/weights'] = 1 378 | vars_corresp['osvos/conv3_3_16/biases'] = 2 379 | vars_corresp['osvos/conv4_3_16/weights'] = 1 380 | vars_corresp['osvos/conv4_3_16/biases'] = 2 381 | vars_corresp['osvos/conv5_3_16/weights'] = 1 382 | vars_corresp['osvos/conv5_3_16/biases'] = 2 383 | 384 | vars_corresp['osvos/score-dsn_2/weights'] = 0.1 385 | vars_corresp['osvos/score-dsn_2/biases'] = 0.2 386 | vars_corresp['osvos/score-dsn_3/weights'] = 0.1 387 | vars_corresp['osvos/score-dsn_3/biases'] = 0.2 388 | vars_corresp['osvos/score-dsn_4/weights'] = 0.1 389 | vars_corresp['osvos/score-dsn_4/biases'] = 0.2 390 | vars_corresp['osvos/score-dsn_5/weights'] = 0.1 391 | vars_corresp['osvos/score-dsn_5/biases'] = 0.2 392 | 393 | vars_corresp['osvos/upscore-fuse/weights'] = 0.01 394 | vars_corresp['osvos/upscore-fuse/biases'] = 0.02 395 | return vars_corresp 396 | 397 | 398 | def _train(dataset, initial_ckpt, supervison, learning_rate, logs_path, max_training_iters, save_step, display_step, 399 | global_step, iter_mean_grad=1, batch_size=1, momentum=0.9, resume_training=False, config=None, finetune=1, 400 | test_image_path=None, ckpt_name="osvos"): 401 | """Train OSVOS 402 | Args: 403 | dataset: Reference to a Dataset object instance 404 | initial_ckpt: Path to the checkpoint to initialize the network (May be parent network or pre-trained Imagenet) 405 | supervison: Level of the side outputs supervision: 1-Strong 2-Weak 3-No supervision 406 | learning_rate: Value for the learning rate. It can be a number or an instance to a learning rate object. 407 | logs_path: Path to store the checkpoints 408 | max_training_iters: Number of training iterations 409 | save_step: A checkpoint will be created every save_steps 410 | display_step: Information of the training will be displayed every display_steps 411 | global_step: Reference to a Variable that keeps track of the training steps 412 | iter_mean_grad: Number of gradient computations that are average before updating the weights 413 | batch_size: Size of the training batch 414 | momentum: Value of the momentum parameter for the Momentum optimizer 415 | resume_training: Boolean to try to restore from a previous checkpoint (True) or not (False) 416 | config: Reference to a Configuration object used in the creation of a Session 417 | finetune: Use to select the type of training, 0 for the parent network and 1 for finetunning 418 | test_image_path: If image path provided, every save_step the result of the network with this image is stored 419 | Returns: 420 | """ 421 | model_name = os.path.join(logs_path, ckpt_name+".ckpt") 422 | if config is None: 423 | config = tf.ConfigProto() 424 | config.gpu_options.allow_growth = True 425 | # config.log_device_placement = True 426 | config.allow_soft_placement = True 427 | 428 | tf.logging.set_verbosity(tf.logging.INFO) 429 | 430 | # Prepare the input data 431 | input_image = tf.placeholder(tf.float32, [batch_size, None, None, 3]) 432 | input_label = tf.placeholder(tf.float32, [batch_size, None, None, 1]) 433 | 434 | # Create the network 435 | with slim.arg_scope(osvos_arg_scope()): 436 | net, end_points = osvos(input_image) 437 | 438 | # Initialize weights from pre-trained model 439 | if finetune == 0: 440 | init_weights = load_vgg_imagenet(initial_ckpt) 441 | 442 | # Define loss 443 | with tf.name_scope('losses'): 444 | if supervison == 1 or supervison == 2: 445 | dsn_2_loss = class_balanced_cross_entropy_loss(end_points['osvos/score-dsn_2-cr'], input_label) 446 | tf.summary.scalar('dsn_2_loss', dsn_2_loss) 447 | dsn_3_loss = class_balanced_cross_entropy_loss(end_points['osvos/score-dsn_3-cr'], input_label) 448 | tf.summary.scalar('dsn_3_loss', dsn_3_loss) 449 | dsn_4_loss = class_balanced_cross_entropy_loss(end_points['osvos/score-dsn_4-cr'], input_label) 450 | tf.summary.scalar('dsn_4_loss', dsn_4_loss) 451 | dsn_5_loss = class_balanced_cross_entropy_loss(end_points['osvos/score-dsn_5-cr'], input_label) 452 | tf.summary.scalar('dsn_5_loss', dsn_5_loss) 453 | 454 | main_loss = class_balanced_cross_entropy_loss(net, input_label) 455 | tf.summary.scalar('main_loss', main_loss) 456 | 457 | if supervison == 1: 458 | output_loss = dsn_2_loss + dsn_3_loss + dsn_4_loss + dsn_5_loss + main_loss 459 | elif supervison == 2: 460 | output_loss = 0.5 * dsn_2_loss + 0.5 * dsn_3_loss + 0.5 * dsn_4_loss + 0.5 * dsn_5_loss + main_loss 461 | elif supervison == 3: 462 | output_loss = main_loss 463 | else: 464 | sys.exit('Incorrect supervision id, select 1 for supervision of the side outputs, 2 for weak supervision ' 465 | 'of the side outputs and 3 for no supervision of the side outputs') 466 | total_loss = output_loss + tf.add_n(tf.losses.get_regularization_losses()) 467 | tf.summary.scalar('total_loss', total_loss) 468 | 469 | # Define optimization method 470 | with tf.name_scope('optimization'): 471 | tf.summary.scalar('learning_rate', learning_rate) 472 | optimizer = tf.train.MomentumOptimizer(learning_rate, momentum) 473 | grads_and_vars = optimizer.compute_gradients(total_loss) 474 | with tf.name_scope('grad_accumulator'): 475 | grad_accumulator = {} 476 | for ind in range(0, len(grads_and_vars)): 477 | if grads_and_vars[ind][0] is not None: 478 | grad_accumulator[ind] = tf.ConditionalAccumulator(grads_and_vars[ind][0].dtype) 479 | with tf.name_scope('apply_gradient'): 480 | layer_lr = parameter_lr() 481 | grad_accumulator_ops = [] 482 | for var_ind, grad_acc in six.iteritems(grad_accumulator): 483 | var_name = str(grads_and_vars[var_ind][1].name).split(':')[0] 484 | var_grad = grads_and_vars[var_ind][0] 485 | grad_accumulator_ops.append(grad_acc.apply_grad(var_grad * layer_lr[var_name], 486 | local_step=global_step)) 487 | with tf.name_scope('take_gradients'): 488 | mean_grads_and_vars = [] 489 | for var_ind, grad_acc in six.iteritems(grad_accumulator): 490 | mean_grads_and_vars.append( 491 | (grad_acc.take_grad(iter_mean_grad), grads_and_vars[var_ind][1])) 492 | apply_gradient_op = optimizer.apply_gradients(mean_grads_and_vars, global_step=global_step) 493 | # Log training info 494 | merged_summary_op = tf.summary.merge_all() 495 | 496 | # Log evolution of test image 497 | if test_image_path is not None: 498 | probabilities = tf.nn.sigmoid(net) 499 | img_summary = tf.summary.image("Output probabilities", probabilities, max_outputs=1) 500 | # Initialize variables 501 | init = tf.global_variables_initializer() 502 | 503 | # Create objects to record timing and memory of the graph execution 504 | # run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE) # Option in the session options=run_options 505 | # run_metadata = tf.RunMetadata() # Option in the session run_metadata=run_metadata 506 | # summary_writer.add_run_metadata(run_metadata, 'step%d' % i) 507 | with tf.Session(config=config) as sess: 508 | print('Init variable') 509 | sess.run(init) 510 | 511 | # op to write logs to Tensorboard 512 | summary_writer = tf.summary.FileWriter(logs_path, graph=tf.get_default_graph()) 513 | 514 | # Create saver to manage checkpoints 515 | saver = tf.train.Saver(max_to_keep=None) 516 | 517 | last_ckpt_path = tf.train.latest_checkpoint(logs_path) 518 | if last_ckpt_path is not None and resume_training: 519 | # Load last checkpoint 520 | print('Initializing from previous checkpoint...') 521 | saver.restore(sess, last_ckpt_path) 522 | step = global_step.eval() + 1 523 | else: 524 | # Load pre-trained model 525 | if finetune == 0: 526 | print('Initializing from pre-trained imagenet model...') 527 | init_weights(sess) 528 | else: 529 | print('Initializing from specified pre-trained model...') 530 | # init_weights(sess) 531 | var_list = [] 532 | for var in tf.global_variables(): 533 | var_type = var.name.split('/')[-1] 534 | if 'weights' in var_type or 'bias' in var_type: 535 | var_list.append(var) 536 | saver_res = tf.train.Saver(var_list=var_list) 537 | saver_res.restore(sess, initial_ckpt) 538 | step = 1 539 | sess.run(interp_surgery(tf.global_variables())) 540 | print('Weights initialized') 541 | 542 | print('Start training') 543 | while step < max_training_iters + 1: 544 | # Average the gradient 545 | for _ in range(0, iter_mean_grad): 546 | batch_image, batch_label = dataset.next_batch(batch_size, 'train') 547 | image = preprocess_img(batch_image[0]) 548 | label = preprocess_labels(batch_label[0]) 549 | run_res = sess.run([total_loss, merged_summary_op] + grad_accumulator_ops, 550 | feed_dict={input_image: image, input_label: label}) 551 | batch_loss = run_res[0] 552 | summary = run_res[1] 553 | 554 | # Apply the gradients 555 | sess.run(apply_gradient_op) # Momentum updates here its statistics 556 | 557 | # Save summary reports 558 | summary_writer.add_summary(summary, step) 559 | 560 | # Display training status 561 | if step % display_step == 0: 562 | print("{} Iter {}: Training Loss = {:.4f}".format(datetime.now(), step, batch_loss), file=sys.stderr) 563 | 564 | # Save a checkpoint 565 | if step % save_step == 0: 566 | if test_image_path is not None: 567 | curr_output = sess.run(img_summary, feed_dict={input_image: preprocess_img(test_image_path)}) 568 | summary_writer.add_summary(curr_output, step) 569 | save_path = saver.save(sess, model_name, global_step=global_step) 570 | print("Model saved in file: %s" % save_path) 571 | 572 | step += 1 573 | 574 | if (step - 1) % save_step != 0: 575 | save_path = saver.save(sess, model_name, global_step=global_step) 576 | print("Model saved in file: %s" % save_path) 577 | 578 | print('Finished training.') 579 | 580 | 581 | def train_parent(dataset, initial_ckpt, supervison, learning_rate, logs_path, max_training_iters, save_step, 582 | display_step, global_step, iter_mean_grad=1, batch_size=1, momentum=0.9, resume_training=False, 583 | config=None, test_image_path=None, ckpt_name="osvos"): 584 | """Train OSVOS parent network 585 | Args: 586 | See _train() 587 | Returns: 588 | """ 589 | finetune = 0 590 | _train(dataset, initial_ckpt, supervison, learning_rate, logs_path, max_training_iters, save_step, display_step, 591 | global_step, iter_mean_grad, batch_size, momentum, resume_training, config, finetune, test_image_path, 592 | ckpt_name) 593 | 594 | 595 | def train_finetune(dataset, initial_ckpt, supervison, learning_rate, logs_path, max_training_iters, save_step, 596 | display_step, global_step, iter_mean_grad=1, batch_size=1, momentum=0.9, resume_training=False, 597 | config=None, test_image_path=None, ckpt_name="osvos"): 598 | """Finetune OSVOS 599 | Args: 600 | See _train() 601 | Returns: 602 | """ 603 | finetune = 1 604 | _train(dataset, initial_ckpt, supervison, learning_rate, logs_path, max_training_iters, save_step, display_step, 605 | global_step, iter_mean_grad, batch_size, momentum, resume_training, config, finetune, test_image_path, 606 | ckpt_name) 607 | 608 | 609 | def test(dataset, checkpoint_file, result_path, config=None): 610 | """Test one sequence 611 | Args: 612 | dataset: Reference to a Dataset object instance 613 | checkpoint_path: Path of the checkpoint to use for the evaluation 614 | result_path: Path to save the output images 615 | config: Reference to a Configuration object used in the creation of a Session 616 | Returns: 617 | """ 618 | if config is None: 619 | config = tf.ConfigProto() 620 | config.gpu_options.allow_growth = True 621 | # config.log_device_placement = True 622 | config.allow_soft_placement = True 623 | tf.logging.set_verbosity(tf.logging.INFO) 624 | 625 | # Input data 626 | batch_size = 1 627 | input_image = tf.placeholder(tf.float32, [batch_size, None, None, 3]) 628 | 629 | # Create the cnn 630 | with slim.arg_scope(osvos_arg_scope()): 631 | net, end_points = osvos(input_image) 632 | probabilities = tf.nn.sigmoid(net) 633 | global_step = tf.Variable(0, name='global_step', trainable=False) 634 | 635 | # Create a saver to load the network 636 | saver = tf.train.Saver([v for v in tf.global_variables() if '-up' not in v.name and '-cr' not in v.name]) 637 | 638 | with tf.Session(config=config) as sess: 639 | sess.run(tf.global_variables_initializer()) 640 | sess.run(interp_surgery(tf.global_variables())) 641 | saver.restore(sess, checkpoint_file) 642 | if not os.path.exists(result_path): 643 | os.makedirs(result_path) 644 | for frame in range(0, dataset.get_test_size()): 645 | img, curr_img = dataset.next_batch(batch_size, 'test') 646 | curr_frame_orig_name = os.path.split(curr_img[0])[1] 647 | curr_frame = os.path.splitext(curr_frame_orig_name)[0] + '.png' 648 | image = preprocess_img(img[0]) 649 | res = sess.run(probabilities, feed_dict={input_image: image}) 650 | res_np = res.astype(np.float32)[0, :, :, 0] > 162.0/255.0 651 | scipy.misc.imsave(os.path.join(result_path, curr_frame), res_np.astype(np.float32)) 652 | print('Saving ' + os.path.join(result_path, curr_frame)) 653 | 654 | 655 | 656 | -------------------------------------------------------------------------------- /osvos_demo.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | """ 3 | Sergi Caelles (scaelles@vision.ee.ethz.ch) 4 | 5 | This file is part of the OSVOS paper presented in: 6 | Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, Luc Van Gool 7 | One-Shot Video Object Segmentation 8 | CVPR 2017 9 | Please consider citing the paper if you use this code. 10 | """ 11 | import os 12 | import sys 13 | from PIL import Image 14 | import numpy as np 15 | import tensorflow as tf 16 | slim = tf.contrib.slim 17 | import matplotlib.pyplot as plt 18 | # Import OSVOS files 19 | root_folder = os.path.dirname(os.path.realpath(__file__)) 20 | sys.path.append(os.path.abspath(root_folder)) 21 | import osvos 22 | from dataset import Dataset 23 | os.chdir(root_folder) 24 | 25 | # User defined parameters 26 | seq_name = "car-shadow" 27 | gpu_id = 0 28 | train_model = True 29 | result_path = os.path.join('DAVIS', 'Results', 'Segmentations', '480p', 'OSVOS', seq_name) 30 | 31 | # Train parameters 32 | parent_path = os.path.join('models', 'OSVOS_parent', 'OSVOS_parent.ckpt-50000') 33 | logs_path = os.path.join('models', seq_name) 34 | max_training_iters = 500 35 | 36 | # Define Dataset 37 | test_frames = sorted(os.listdir(os.path.join('DAVIS', 'JPEGImages', '480p', seq_name))) 38 | test_imgs = [os.path.join('DAVIS', 'JPEGImages', '480p', seq_name, frame) for frame in test_frames] 39 | if train_model: 40 | train_imgs = [os.path.join('DAVIS', 'JPEGImages', '480p', seq_name, '00000.jpg')+' '+ 41 | os.path.join('DAVIS', 'Annotations', '480p', seq_name, '00000.png')] 42 | dataset = Dataset(train_imgs, test_imgs, './', data_aug=True) 43 | else: 44 | dataset = Dataset(None, test_imgs, './') 45 | 46 | # Train the network 47 | if train_model: 48 | # More training parameters 49 | learning_rate = 1e-8 50 | save_step = max_training_iters 51 | side_supervision = 3 52 | display_step = 10 53 | with tf.Graph().as_default(): 54 | with tf.device('/gpu:' + str(gpu_id)): 55 | global_step = tf.Variable(0, name='global_step', trainable=False) 56 | osvos.train_finetune(dataset, parent_path, side_supervision, learning_rate, logs_path, max_training_iters, 57 | save_step, display_step, global_step, iter_mean_grad=1, ckpt_name=seq_name) 58 | 59 | # Test the network 60 | with tf.Graph().as_default(): 61 | with tf.device('/gpu:' + str(gpu_id)): 62 | checkpoint_path = os.path.join('models', seq_name, seq_name+'.ckpt-'+str(max_training_iters)) 63 | osvos.test(dataset, checkpoint_path, result_path) 64 | 65 | # Show results 66 | overlay_color = [255, 0, 0] 67 | transparency = 0.6 68 | plt.ion() 69 | for img_p in test_frames: 70 | frame_num = img_p.split('.')[0] 71 | img = np.array(Image.open(os.path.join('DAVIS', 'JPEGImages', '480p', seq_name, img_p))) 72 | mask = np.array(Image.open(os.path.join(result_path, frame_num+'.png'))) 73 | mask = mask//np.max(mask) 74 | im_over = np.ndarray(img.shape) 75 | im_over[:, :, 0] = (1 - mask) * img[:, :, 0] + mask * (overlay_color[0]*transparency + (1-transparency)*img[:, :, 0]) 76 | im_over[:, :, 1] = (1 - mask) * img[:, :, 1] + mask * (overlay_color[1]*transparency + (1-transparency)*img[:, :, 1]) 77 | im_over[:, :, 2] = (1 - mask) * img[:, :, 2] + mask * (overlay_color[2]*transparency + (1-transparency)*img[:, :, 2]) 78 | plt.imshow(im_over.astype(np.uint8)) 79 | plt.axis('off') 80 | plt.show() 81 | plt.pause(0.01) 82 | plt.clf() 83 | -------------------------------------------------------------------------------- /osvos_parent_demo.py: -------------------------------------------------------------------------------- 1 | """ 2 | Sergi Caelles (scaelles@vision.ee.ethz.ch) 3 | 4 | This file is part of the OSVOS paper presented in: 5 | Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, Luc Van Gool 6 | One-Shot Video Object Segmentation 7 | CVPR 2017 8 | Please consider citing the paper if you use this code. 9 | """ 10 | import os 11 | import sys 12 | import tensorflow as tf 13 | slim = tf.contrib.slim 14 | # Import OSVOS files 15 | root_folder = os.path.dirname(os.path.realpath(__file__)) 16 | sys.path.append(os.path.abspath(root_folder)) 17 | import osvos 18 | from dataset import Dataset 19 | 20 | # User defined parameters 21 | gpu_id = 0 22 | 23 | # Training parameters 24 | imagenet_ckpt = 'models/vgg_16.ckpt' 25 | logs_path = os.path.join(root_folder, 'models', 'OSVOS_parent') 26 | store_memory = True 27 | data_aug = True 28 | iter_mean_grad = 10 29 | max_training_iters_1 = 15000 30 | max_training_iters_2 = 30000 31 | max_training_iters_3 = 50000 32 | save_step = 5000 33 | test_image = None 34 | display_step = 100 35 | ini_learning_rate = 1e-8 36 | boundaries = [10000, 15000, 25000, 30000, 40000] 37 | values = [ini_learning_rate, ini_learning_rate * 0.1, ini_learning_rate, ini_learning_rate * 0.1, ini_learning_rate, 38 | ini_learning_rate * 0.1] 39 | 40 | # Define Dataset 41 | train_file = 'train_parent.txt' 42 | dataset = Dataset(train_file, None, './DAVIS', store_memory=store_memory, data_aug=data_aug) 43 | 44 | # Train the network 45 | with tf.Graph().as_default(): 46 | with tf.device('/gpu:' + str(gpu_id)): 47 | global_step = tf.Variable(0, name='global_step', trainable=False) 48 | learning_rate = tf.train.piecewise_constant(global_step, boundaries, values) 49 | osvos.train_parent(dataset, imagenet_ckpt, 1, learning_rate, logs_path, max_training_iters_1, save_step, 50 | display_step, global_step, iter_mean_grad=iter_mean_grad, test_image_path=test_image, 51 | ckpt_name='OSVOS_parent') 52 | 53 | with tf.Graph().as_default(): 54 | with tf.device('/gpu:' + str(gpu_id)): 55 | global_step = tf.Variable(max_training_iters_1, name='global_step', trainable=False) 56 | learning_rate = tf.train.piecewise_constant(global_step, boundaries, values) 57 | osvos.train_parent(dataset, imagenet_ckpt, 2, learning_rate, logs_path, max_training_iters_2, save_step, 58 | display_step, global_step, iter_mean_grad=iter_mean_grad, resume_training=True, 59 | test_image_path=test_image, ckpt_name='OSVOS_parent') 60 | 61 | with tf.Graph().as_default(): 62 | with tf.device('/gpu:' + str(gpu_id)): 63 | global_step = tf.Variable(max_training_iters_2, name='global_step', trainable=False) 64 | learning_rate = tf.train.piecewise_constant(global_step, boundaries, values) 65 | osvos.train_parent(dataset, imagenet_ckpt, 3, learning_rate, logs_path, max_training_iters_3, save_step, 66 | display_step, global_step, iter_mean_grad=iter_mean_grad, resume_training=True, 67 | test_image_path=test_image, ckpt_name='OSVOS_parent') 68 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | tensorflow_gpu==1.1.0 2 | scipy==0.18.1 3 | matplotlib==1.5.3 4 | numpy==1.12.1 5 | Pillow==4.1.1 6 | tensorflow==1.1.0 7 | --------------------------------------------------------------------------------