libuyu

Actually, the results on Pascal Voc have been listed on the poster of my work, and the AP of Focal Loss and GHM are 74.5 vs 74.8. And for training...

I can share with you my slides for the oral presentation, it has more details than the poster. And the results on VOC are on page 10. https://drive.google.com/open?id=1H3tfg2d3NdLPQ7HHtZSgecE7ybN7Z-uS

@fengxiuyaun could you provide more details? Such as the location where the error appear, and whether you write a new data_loader (since the default coco data set code do not...

Well done! So if you use binary classification (one class det), just modify this line. It seems that mmdetection hasn't consider one class output channel now. I will talk with...

You can refer to this closed issue: https://github.com/libuyu/GHM_Detection/issues/2

If there is no region to ignore, your mask should be 1 everywhere. And each pixel in the final predicted map is a sample, and the pos/neg samples are not...

@sjtuliuqin Yes, you got it. And for the theory of the "acc_sum", you can refer to the part "EMA" on page 4 in our paper https://arxiv.org/pdf/1811.05181.pdf. Thank you for your...

If applied on softmax cross entropy loss, the form needs modification. We haven't studied this yet. And if you have any good idea, welcome to discuss in mail. :)

@xialuxi hello, M is the "n" in https://github.com/libuyu/GHM_Detection/blob/0b3917d95870382b1e01c9917e042a93436c5e32/mmdetection/mmdet/core/loss/ghm_loss.py#L41 and https://github.com/libuyu/GHM_Detection/blob/0b3917d95870382b1e01c9917e042a93436c5e32/mmdetection/mmdet/core/loss/ghm_loss.py#L54 Sorry for the different notation.

Since the equation has the form sum[i+1] = mmt * sum[i] + (1 - mmt) * num[i], momentum should have the range [0, 1). So if momentum = 1.0, the...