about the result。

[lw19951231]

I found that you had a low result in the previous question（https://github.com/bellos1203/STPN/issues/4）, but then deleted the question. Did you finally find the reason? thx，：）

[demianzhang]

Yes，but I do not have a try. I ask the issue 2’s author about his method. I find differences between his and mine. I will have a try so I delete it. The only difference is the order between sample 10fps and extract flow. He used opencv to downsample and make the videos into 10fps and then dense_flow. I dense_flow and then choose one image in every three images. Other operation is the same.

1. sample 10fps
2. keep ratio 340x180 and resize the small edge to 256
3. extract flow（dense_flow）
4. center crop 224x224, each 16 frames are sent to i3d

I wish this will help you.

[lw19951231]

Thanks for your prompt reply. my approach is the same as yours, two different methods may affect optical flow extraction? I'll try your method, thx. : )

[demianzhang]

Please remember to sample 10fps and resize then extract flow.

[Rheelt]

Hi, Thank you for your advice! First, i use "ffmpeg -i file -q:v 1 -r 10 -s 340*256 /img%05d.jpg" to extract the image. But then I used the script written by Python calling opencv to extract the optical flow information too slowly. (PS: i tried adding this command video_stream.set(CV_CAP_PROP_FPS,10); to dense_flow - dense_flow_gpu.cpp, but it didn't work.) So do you @lw19951231 @demianzhang have any scripts that can extract optical streams using GPU?

[demianzhang]

yjxiong's dense_flow has more introduction of how to use. Please compiled it with opencv2 in cuda8, opencv3 may cause the speed slowly. The input of dense_flow is video, you can sample 10fps and then synthesize them as video.

[Rheelt]

Thank you @demianzhang ! I used the newly extracted features and adopted the training model provided by the author STPN, and the results was mAP@tIoU=0.5 = 0.156. Then I trained the model myself (I didn't change the training configuration), and the results were terrible (mAP@tIoU=0.5 = 0.0001). Do you know why? Thank you.

[demianzhang]

You can analyze the result by only rgb proposal or flow proposal to find the reason. I did not have a try. It seems like feature is well however some problem happens to others.