└── README.md


/README.md:
--------------------------------------------------------------------------------
 1 | # IFDD
 2 | Few-shot NEU-DET (FS-ND)  
 3 | This is the dataset used in our paper Few-Shot Steel Surface Defect Detection (will be published soon). The dataset is reconstructed from the famous [NEU-DET](https://ieeexplore.ieee.org/abstract/document/8709818) dataset and designed specially for few-shot defect detection.  
 4 | The FS-ND can be downloaded [here](https://drive.google.com/drive/folders/1Somtykp_DwqGTe5by9PEfPDxPqYRFu-L)
 5 | If you use the FS-ND, please cite our paper:  
 6 | @article{wang2021few,  
 7 |   title={Few-Shot Steel Surface Defect Detection},  
 8 |   author={Wang, Haohan and Li, Zhuoling and Wang, Haoqian},  
 9 |   journal={IEEE Transactions on Instrumentation and Measurement},  
10 |   year={2021},  
11 |   publisher={IEEE}  
12 | }
13 | 
14 | ## How to use FS-ND?
15 | There are 4 zip files in the link above: **NEU-DET-METATRAIN**, which is used for meta-training (base training), and **5shot1-100**, **10shot1-100** and **30shot1-100**, which are used for meta-testing (few-shot fine-tuning). For k-shot experiment, you should first train your model with **NEU-DET-METATRAIN**, then evaluate it with the corresponding **kshot1-100**, where k=5,10,30.  
16 | There are 100 data folder in the **kshot1-100.zip**: **kshot1**, **kshot2**...**kshot100**. Each **kshoti** (k=5,10,30 and i = 1,2,3...100) folder has the same structure:  
17 | 
18 | **kshoti  
19 | ├── annotations   # two json files, the annoations for train2017 and val2017  
20 | ├── train2017       # support examples (3*k images)  
21 | ├── val2017          # query images, this folder stays same when k or i changes    
22 | ├── trainval.txt     # intermediate result when constructing data, no attention needed  
23 | ├── test.txt          # intermediate result when constructing data, no attention needed**  
24 | The **kshoti** has the same structure as the famous **COCO** dataset, so you can use the same dataloader with minor adjustments. The two txt file can also be useful if you would like to further analysis the data.   
25 | The **NEU-DET-METATRAIN** shares the same structure with **kshoti**, where its **train2017** contains more images and its **val2017** can be utilized to evaluate the pre-trained model. However, the results on **NEU-DET-METATRAIN/val2017** are just intermediate results and do not reflect the few-shot performance.  
26 | 
27 | ## Why we reconstruct NEU-DET?
28 | Although few-shot defect detection technique is quite valuable due to the scarcity of industrial data, there is no publicly available few-shot defect detection dataset. We can utilize the existing NEU-DET to evaluate few-shot detection tasks, but it actually introduces much randomness.  
29 | First, the different training and testing split has an unpredictable effect, even the same model can perform differently in different works.  
30 | Second, the sample of support examples makes the results unstable, as the quality of each training image differs a lot.  
31 | To bridge this gap, we reconstruct NEU-DET into the few-shot form, namely few-shot NEU-DET (FS-ND). 
32 | 
33 | ## How we make FS-ND?
34 | we fix the partition of training and testing data, and for each k-shot setting (k=5,10,30), we sample 100 groups support examples to reduce the randomness.  
35 | Specifially, three types of defects (inclusion, rolled-in scales and scratches) are utilized as the base data for training, and the others (crazing, patches and pitted surface) are reserved for testing.  
36 | For more details about how the NEU-DET is reconstructed, please refer to our paper (coming soon)  
37 | 


--------------------------------------------------------------------------------