search

Megvii-BaseDetection / OTA

Official implementation of our CVPR2021 paper "OTA: Optimal Transport Assignment for Object Detection" in Pytorch.
Apache License 2.0
242 stars 24 forks source link

Reason for using log scale sinkhorn algorithm #17

Closed jihwanp closed 1 year ago

jihwanp commented 1 year ago

Hi Thanks for providing wonderful work! I have a question about the sinkhorn distances implementation.

In the forward function of sinkhorn distances (below), you use log scale updating rule which is not same as original algorithm, and add v_t,u_t to update v_t+1,u_t+1. Whats the reason for that? https://github.com/Megvii-BaseDetection/OTA/blob/2c85b4d0f9031396854aae969330dde2ab5eacbd/playground/detection/coco/ota.x101.dcnv2.fpn.coco.800size.1x/fcos.py#L693-L699

utkutpcgl commented 1 year ago

Thanks for the great contribution and thanks for asking @jihwanp. I am also curious about this. Could anyone find an answer to this issue?