About the image-level split loss

[Yinyf0804]

Thanks for your great work! I have some questions about the image-level split loss.

1. What's difference between Eq.5 and Eq.8, since they look the same in format. Are the differences in the pseudo labels "ti,tj"?
2. The "instance-level split". Does the instance mean each pseudo-ground truth or each proposal/anchor?
3. How are the background proposals defined?

Looking forward to your reply.

[1170300714]

收到，感谢您的来件

[1170300714]

Thanks for your issues.

1. You are right, they look the same in format but focus on different "level". The Eq.5 indicates the "image loss" where it cares about the all positive proposals of this image I, while the Eq.8 indicates the 'instance loss' where it cares about those positive proposals which are assigned with this pseudo box(instance) S. In addition, the ti and tj indicate the indexes of proposals of RPN and RoI head respectively.
2. The "instance" means each pseudo-ground truth.
3. The definition of the background proposals are same with the label assignment of most detection method. Specifically，the IoU between one proposals and pseudo gt box is lower than 0.5.

[Yinyf0804]

Thanks so much for your reply.