search

yjxiong / temporal-segment-networks

Code & Models for Temporal Segment Networks (TSN) in ECCV 2016
BSD 2-Clause "Simplified" License
1.54k stars 477 forks source link

Use other method instead dense_flow #273

Open dagongji10 opened 5 years ago

dagongji10 commented 5 years ago

I hope to use PWC-Net extract optflow because dense-flow is too slow. But when I use PWC-Net optflow to train model, the acc is below 70%. And I also try farneback(t=0) instead tvl1(t=1) when use dense-flow to extract optflow, but model acc is only 75%. The train process followed README file. Is there any train tips can improve acc?

Fovever1 commented 5 years ago

i meet same problem,when i use other way to extract frame, for example,ffmpeg, model acc doesn't get the paper result, it seems like it has the relationship with frame format or flow format,maybe the format is not same as paper,i think it leads this problem.

Fovever1 commented 5 years ago

@dagongji10 Do you find the cause?

dagongji10 commented 5 years ago

@dagongji10 Do you find the cause?

Maybe you need to modify parameter:num_segments. The default value is 3 but best value in paper is 7. In fact, we only need to train flow mode because rgb mode with same input(rgb frame) while optical flow may have great difference. I tried to train flow mode with farneback flow but the results were not as good as the article suggested. Maybe tvl1 is best for TSN.