Vision: Why does the rpn use the L1_Loss?

[TeeyoHuang]

https://github.com/pytorch/vision/blob/master/torchvision/models/detection/rpn.py#L426

the code in the rpn.py , line 426 as follows:

box_loss = F.l1_loss(
pred_bbox_deltas[sampled_pos_inds],
regression_targets[sampled_pos_inds],
reduction="sum",
) / (sampled_inds.numel())

However, as said in the paper of Faster RCNN, the loss funtion used in the rpn training stage is smooth_L1_LOSS.

and I found that when computing the rcnn_box_loss, the loss function used in the torchvsion is Smooth_L1_Loss,:
https://github.com/pytorch/vision/blob/master/torchvision/models/detection/roi_heads.py#L47

Why not use the Smooth_L1_LOSS in both places ?

[fmassa]

Hi,

Great question. It turns out that using smooth_l1_loss actually gave slightly worse results for COCO for detection and instance segmentation, and has also been replaced by l1_loss in detectron2, see https://github.com/facebookresearch/detectron2/blob/61c15d20f804c0393d412fb3467649ea8e50fb57/detectron2/config/defaults.py#L212 and https://github.com/facebookresearch/detectron2/blob/61c15d20f804c0393d412fb3467649ea8e50fb57/detectron2/config/defaults.py#L280 .

For that reason (and also for simplicity because PyTorch smooth_l1_loss doesn't have the beta argument, and Mask R-CNN already has a ton of arguments), we decided to simply use the l1_loss in torchvision.

Let me know what you think.

I'm closing the issue, but let me know if you have further questions.

[TeeyoHuang]

Hi,

Great question. It turns out that using smooth_l1_loss actually gave slightly worse results for COCO for detection and instance segmentation, and has also been replaced by l1_loss in detectron2, see https://github.com/facebookresearch/detectron2/blob/61c15d20f804c0393d412fb3467649ea8e50fb57/detectron2/config/defaults.py#L212 and https://github.com/facebookresearch/detectron2/blob/61c15d20f804c0393d412fb3467649ea8e50fb57/detectron2/config/defaults.py#L280 .

For that reason (and also for simplicity because PyTorch smooth_l1_loss doesn't have the beta argument, and Mask R-CNN already has a ton of arguments), we decided to simply use the l1_loss in torchvision.

Let me know what you think.

I'm closing the issue, but let me know if you have further questions.

Thank you for your reply!
I also found that the smooth_l1_loss could cause slightly worse results and I just want to confirm my idea.
Thank you very much!