Evaluate your method on Yolo5

[Dongjiuqing]

You show some examples of failures in the experiments. Actually, I guess that was caused by the weak feature extractor. There are some latest feature exactors like YOLO-v5, v7, and transformers. I wonder if it will take a lot of time to apply your method to these networks? I want to apply your method to YOLO v5. Could you give me some suggestions? I think you are experienced.

[cxliu0]

Feature extractors could be one of the reasons that limit the detection performance.

In addition, I think the deployment of our approach on YOLOv5 is similar to OA-MIL FasterRCNN (please refer to standard_roi_head_oamil.py for implementation details). It may not take much time to apply OA-MIL to YOLOv5. You could implement it as follows:

1. construct object bags 
   - each positive object bag is composed of anchors that correspond to the same ground-truth box
2. perform instance selection on each object bag, and output the best selected instances (Eq. 4 in the paper)
3. compute instance selection loss for object bags
4. compute classification and localization losses based on the best selected instances

[JiuqingDong]

I read your code. The difference between 'xxxxxx_r50_fpn_coco_oamil.py' and 'xxxxxx_r50_fpn_coco.py' is that you defined two new classes in these two files as follows.

You give me four suggestions in the previous comments, and all of them are implemented in these two files, right? And then, you register these new classes in the 'init' file. Did you change the other source code in mmdetection? (or add new files).

[cxliu0]

The implementations of OA-MIL are in these two files. No other modification is required.

[iu110]

您好，请问您在yolo上实现了作者的方法了吗？

[JiuqingDong]

您好，请问您yolo上实现了作者的方法吗？

没有。我尝试了但没有成功。