Result mismatch in cascade-rcnn

[duanduanduanyuchen]

I use the config here to train the model and found something unusual. The final bbox_mAP is 35.5 but not 40.3 in the table. Then I found that the default pre-trained weight of r50 has changed in the new version. New pre-trained weights have a higher top-1 acc in classification(76.130%->80.858%) (check here) but caused a worse performance in detection(40.3%->35.5%). When I change the pre-trained weight back to the older one, I finally get 40.3% mAP.

I think the default config should change the link of pre-trained weight to adapt to the new version of torchvision.

Does anyone know the reason why the worse performance was caused by the better pre-trained weights?

[shinya7y]

I think the default config should change the link of pre-trained weight to adapt to the new version of torchvision.

The latest mmcv seems to load old weights for backward compatibility.
open-mmlab/mmcv#1867

[duanduanduanyuchen]

Thanks for letting me know that!