jiyanggao/turn-tap

TURN TAP: Temporal Unit Regression Network for Temporal Action Proposals

turn-tap's Issues

Problem in resnet feature for test set

Hi @jiyanggao ,

Recently, when I was trying to use the 'test_fc6_16_overlap0.5_resnet.tar', I found that there is not enough data in this folder, since the number of files in resnet and denseflow folder is not the same. And I checked, there is half of the feature only in resnet test folder.

Could u please share the code of extracting feature or just upload a new version of the feature? I really appreciate it, thanks!

Best,
June

Question about training

Hi, thanks for your helpful code. I try to train my own dataset in this way. However, the loss keep oscillating, I tried different lr, but still. May I ask what the reason is, can you give some other opinion?
Thank you very much!

Any pre-trained model to download ?

Hi, Jiyang. Would you share your trained model so that we can use it for temporal proposal extraction?

denseflow code

hi,thank you for sharing code. the denseflow cnn feature you metioned i can not found. could you give some more details ?

How many hours does it take until the model convergence on your device?

Hi @jiyanggao, thanks for your helpful code. How many hours does it take until the model convergence on your device?

question about denseflow cnn feature

Could you please give me any tips about denseflow cnn feature?I want to test my own video.

about SCNN-localization evaluate the proposal

Hello,Jiyang,When I tried to fed the proposals to SCNN-localization, I meet some problem.
SCNN-loc could not generate some prob file .

questions about the length of clips in test_swin.txt file

Hi, in the test_swin.txt file, I notice that the clips of different videos are of different length.
For example, clips of video_test_0001270 are all of 16 frames, while clips of video_test_0001313 are all of 32 frames, and clips of some other videos are of mixed length of 16 or 32 frames. Could you provide the rules of generating clips during test?

dataset next batch issue.

Hi Jiyang!
I have been followed your a series of work in action detection area. Very solid work.

Thanks to open source your TURN-TAP model.

When reading your code, I found that in dataset.py, you give mini-batch by using:

The point is: random sample here is not a good choice. when training, some samples may be trained multiple times while others may not be trained. That's werid.

In other words, you cannot train a whole epoch.

I think pytorch's dataset wrapper would be a good choice.

Post-processing Issues

I have viewed your code in post-processing and have the following questions.

First of all, in your code, you computed the prob distribution based on clip length, and, thus, shorter clip will be given greater probability due to its massive quantity in clip generation. In this case, you punished the proposals with longer duration, but I don't the reasons.

Second, I have not found NMS code in post-processing according to your paper.