How to determine the boundary area of the object of interest?

[YH-2023]

1.How to determine the boundary area of the object of interest? 2.The paper mentions: the class-agnostic activation maps can be used to extract class-agnostic object bounding boxes for localization. How to locate the code part that is not found?

[Sierkinhane]

Hi, here are the codes for extracting object bounding boxes.

[YH-2023]

@Sierkinhane Hi ，

1. Can I not use the bounding box extraction method mentioned in the code, but can I only perform conventional target detection, classification and regression operations on the foreground features? Does this apply to ordinary target detection? In other words， If I want to use CCAM for target detection, should I use the separated foreground features for classification and regression, or should I use the obtained object boundary area for classification and regression?
2. Can the CCAM algorithm locate all instance objects in the image, or only the known objects in the training set?
3. I have a question about the code. Is the test in it used for verification? https://github.com/CVI-SZU/CCAM/blob/78448d4654465251dfb0b30ba4d7f82af84b0919/WSOL/train_CCAM_CUB.py#L131 Thanks

[Sierkinhane]

Unfortunately, the proposed method currently only works well on foreground-background separation. We tried to use the learned features for classification but it didn't get better results than traditional classification methods. We argue the model cannot learn discrimative features via the proposed foreground-background contrastive loss. Maybe you can try to involve more activation channels and build additional contrastive losses upon the foreground-background contrast.

Exactly, it is for evalution.