Question about the reslute mAP.

[ayumiymk]

Does anyone get the comparable mAP on VOC07 in the origin paper?

[zhanghang1989]

The author reported here: https://github.com/longcw/faster_rcnn_pytorch#training-on-pascal-voc-2007

[ayumiymk]

But the author got a 0.661 mAP on VOC07 while the origin paper got a 0.699 mAP. The author said we may need to tune the loss function defined in faster_rcnn/faster_rcnn.py by ourself. So what I want to know is the config to get a mAP comparable to the origin paper.

[acgtyrant]

I rewrite the project, and calculate its mAP, only 0.356.

[zchrissirhcz]

I just git cloned this project and train on my 1080Ti machine. The finally mAP@100000iteration is 65.6%.

[brisker]

@zchrissirhcz @longcw so what does the author mean by "tune the loss by ourselves in faster_rcnn.py" to get 0.699 mAP?

[zchrissirhcz]

@brisker In my opinion, I appreciate longcw's re-implementation. However, I'd like to use implementations with equal or higher accuracy than the paper. I suggest try this implementation:https://github.com/ruotianluo/pytorch-faster-rcnn

[brisker]

@zchrissirhcz https://github.com/ruotianluo/pytorch-faster-rcnn this implementation seems to be more consistent with rbg's original caffe version implementation. But I like longcw's implementation more. But still, I have no idea where to modify the loss function, as longcw have indicated. Do you know?