Zhang Yifeng

The accuracy i test is lower than that in the readme file

1

I try to evaluate the accuracy of mobilenet v2 using the caffemodel & prototxt in the git, however the result of top1 is 0.6977 in the validation set

x1, y1, x2, y2, x3, y3, x4, y4 are 8 variances, and the former the xmin, ymin, xmax, ymax can be computed by these 8 variances, so why 12 variances...

All the boxes are calculated in focal loss, function MineHardExamples should be modified to cover all boxes. Another tricky way is set _pos_neg_ratio_ into a very large number, but takes...

In the appendix E.2, the author says using lambda layer in c4 and c5 stage is good enough for the result, while keeping the speed not too slow.

Signle decode cost 137 ms per image Batch decode cost 0.64 ms per image I guess the reason is the torchnvjpeg.Decoder parameter different, so I set the param without torch.cuda.current_stream(device_id),...