search

open-mmlab / mmdetection

OpenMMLab Detection Toolbox and Benchmark
https://mmdetection.readthedocs.io
Apache License 2.0
29.71k stars 9.48k forks source link

RTMDet different loss than official paper ? #11732

Open PushpakBhoge opened 6 months ago

PushpakBhoge commented 6 months ago

The official paper of RTM Det https://arxiv.org/pdf/2212.07784 link the RTMDet of mmdet as official implementation of the paper but there is one weird thing I have noticed that official paper mentions that the bbox_loss is -log(IOU)

but in config it using GIOU loss ? I thought maybe in code this log was being taken but I did not find any such operations in class RTMDetHead method ?

is there any reason why this part is different from the paper ? Also why have GIoU loss and not CIoU loss

tagging authors here thank you for your help @RangiLyu @ZwwWayne

tinggh commented 2 weeks ago

The cls loss is quality focal loss just as the gfl paper, but not as the offical paper. Also the region cost is not uesed inthe offical code.

tinggh commented 2 weeks ago

The cls loss is quality focal loss just as the gfl paper, but not as the offical paper. Also the region cost is not uesed inthe offical code.

oh, I know the loss in the offical paper is just used as the cost in dynamic soft label assigner, the head cls loss and box loss is just as the quality focal loss and giou loss.

PushpakBhoge commented 2 weeks ago

yes you are right I found that piece of code in the assigned class Thank you for your help @tinggh