Why do you use 4ConvFC head instead of the FC head in FPN?

[gugite]

Thank you for your excellent work! I notice that you use the 4ConvFC head instead of the standard FC head. As far as I know, introducing 4 Conv will give some AP gains in terms of the object detector. Do you have the results with the standard FC head? Thanks.

https://github.com/wusize/ovdet/blob/cbeb3d40492b7c871e454de32039298c475f5ac1/configs/_base_/models/faster-rcnn_r50_fpn_syncbn.py#L14

[wusize]

Hi, Thanks for your interest in our work! Yes, I used 4ConvFC to obtain a stronger object detector. This was for fair comparison with previsous papers that usually adopted a strong baseline, e.g. ViLD, DetPro. However, there is also a config file using a simple c4 based detector, which yields good result on novel AP and of course much lower base AP compared to the FPN+4ConvFC based methods.

[gugite]

Thank you for your reply! I notice that you use a 0.02*2 learning rate for FPN. Is that for 16 or 32 batch-size?

https://github.com/wusize/ovdet/blob/652da9904d718b8cd4aed1261171533b7ef4eb58/configs/baron/ov_coco/faster_rcnn_r50_fpn_syncbn_90kx2.py#L71

Moreover, could you please tell me when will release the pre-trained model? Thank you very much!

[wusize]

Yes, that's because I use 16 gpus (batchsize = 16x2=32) to train it. The method is initially developed on detecron2. Currently we are having problem on the training speed and accuracy when implementing it on MMDet3.x. The models trained on COCO are expected to be released this weekend. I will notify you once they are ready.

[gugite]

Thank you so much！

[wusize]

Hi, the testing and training on coco are supported now.