Scalsol / mega.pytorch

Memory Enhanced Global-Local Aggregation for Video Object Detection, CVPR2020
Other
570 stars 116 forks source link

About reproducing FGFA results #103

Closed HzZHoO closed 3 years ago

HzZHoO commented 3 years ago

Thanks again for your great work! I have a question about the mAP result of FGFA.

I did two changes for FGFA: allowing one GPU hold more than one training sample, training with current feature aggregated (which is good for the final result). The question is, following the default training schedule of this project (batchsize=8, lr=0.02, (40000, 20000)iters...), but I cannot reach your result, which is 78.0. My result is 76.85 and official result in paper with different training schedule is 76.3.

The gap is not marginal. Is this because I miss some tricks or training iterations is not enough?

Looking forward to your help!