Question about the result of FsDetView

[Wei-i]

Hi, thanks for your great work and source code! I also want to know why the mAP of FsDetView * drops a lot in your paper, while in FsDetView's paper, on 10-shot COCO setting, it can achieve 12.5. Is it the same as the difference of Meta R-CNN configuration? I would be grateful if you could apply for me.

[Wei-i]

Another issue I met was that when I run python test_net.py --num-gpus 8 MODEL.WEIGHTS log/model_redetect.pth without --double-rpn args to test for performance. However from the screen, it seems that the parameters of finetuned_objectness_logits didn't use?

--log [08/13 20:30:22 c2.checkpoint.checkpoint]: The checkpoint contains parameters not used by the model: proposal_generator.rpn_head.finetuned_objectness_logits.{weight, bias}

[zb1439]

Hi, thanks for your great work and source code! I also want to know why the mAP of FsDetView * drops a lot in your paper, while in FsDetView's paper, on 10-shot COCO setting, it can achieve 12.5. Is it the same as the difference of Meta R-CNN configuration? I would be grateful if you could apply for me.

Yes. The reason is the same as https://github.com/Megvii-BaseDetection/GFSD/issues/1

[zb1439]

Another issue I met was that when I run python test_net.py --num-gpus 8 MODEL.WEIGHTS log/model_redetect.pth without --double-rpn args to test for performance. However from the screen, it seems that the parameters of finetuned_objectness_logits didn't use?

--log [08/13 20:30:22 c2.checkpoint.checkpoint]: The checkpoint contains parameters not used by the model: proposal_generator.rpn_head.finetuned_objectness_logits.{weight, bias}

The model will use a general RPN instead of the bias-balanced RPN and load the pretrained RPN's weight if the double-rpn option is disabled.

[Wei-i]

Another issue I met was that when I run python test_net.py --num-gpus 8 MODEL.WEIGHTS log/model_redetect.pth without --double-rpn args to test for performance. However from the screen, it seems that the parameters of finetuned_objectness_logits didn't use? --log [08/13 20:30:22 c2.checkpoint.checkpoint]: The checkpoint contains parameters not used by the model: proposal_generator.rpn_head.finetuned_objectness_logits.{weight, bias}

The model will use a general RPN instead of the bias-balanced RPN and load the pretrained RPN's weight if the double-rpn option is disabled.

Thanks! I am confused if the bias-balanced RPN is only used for inference ?

[zb1439]

Another issue I met was that when I run python test_net.py --num-gpus 8 MODEL.WEIGHTS log/model_redetect.pth without --double-rpn args to test for performance. However from the screen, it seems that the parameters of finetuned_objectness_logits didn't use? --log [08/13 20:30:22 c2.checkpoint.checkpoint]: The checkpoint contains parameters not used by the model: proposal_generator.rpn_head.finetuned_objectness_logits.{weight, bias}

The model will use a general RPN instead of the bias-balanced RPN and load the pretrained RPN's weight if the double-rpn option is disabled.

Thanks! I am confused if the bias-balanced RPN is only used for inference ?

Yes, this is mentioned in the implementation details in the supplementary materials.

[Wei-i]

Another issue I met was that when I run python test_net.py --num-gpus 8 MODEL.WEIGHTS log/model_redetect.pth without --double-rpn args to test for performance. However from the screen, it seems that the parameters of finetuned_objectness_logits didn't use? --log [08/13 20:30:22 c2.checkpoint.checkpoint]: The checkpoint contains parameters not used by the model: proposal_generator.rpn_head.finetuned_objectness_logits.{weight, bias}

The model will use a general RPN instead of the bias-balanced RPN and load the pretrained RPN's weight if the double-rpn option is disabled.

Thanks! I am confused if the bias-balanced RPN is only used for inference ?

Yes, this is mentioned in the implementation details in the supplementary materials.

Thanks！ I want to know how to read your supplementary materials. Could you give me a link ?

[zb1439]

The link in readme should be ok.

[Wei-i]

@zb1439 I am sorry to disturb you again, do you have the version of detectron2's implementation?

[zb1439]

again

Sorry but we do not have a d2 version code. I feel like some minor but careful modifications on config should be almost sufficient for transferring the code. Also please feel free to question.

[Wei-i]

again

Sorry but we do not have a d2 version code. I feel like some minor but careful modifications on config should be almost sufficient for transferring the code. Also please feel free to question.

Thanks again for your selfless help :)))