Trained Weights are present on the link. The weights of best performing model are present in the exp120/ folder.
Reproducability of Best Perfoming Model:
- Resizing - Assuming we have input annotated dataset with mamogram 4k images, first we need to resize the images to 1k & 2k resolution using the Pytorch/OpenCv standard utils with Bilinear Interpolation. Use this resize script to corresponding resize groundtruth annotations.
- Systematic Crop - We crop images of 2k and 4k in fragments of size 1k and corresponding create annotations for each fragment. Use this crop script, You might need to change the path to your own dataset.
- Training the Model - Configure the file coco128.yml to add paths for the datasets, it should include 1k images and 1k crops from 2k & 4k images. Run train.py
- Testing at three scales - Test the model obtained in the previous step on 1k, 2k & 4k resolution datasets respectively using detect.py.
- Merging the obtained predictions - Merge the predictions obtained in the previous step using the script merge_wbf.py
- Evaluation - To get the values FPI at various Sensitivities use FROC.py
Credits:
- Used the base implementation of Yolov5 from https://github.com/ultralytics/yolov5
- Wbf script for merging is taken from https://github.com/ZFTurbo/Weighted-Boxes-Fusion