search

MohsenFayyaz89 / SCT

SCT: Set Constrained Temporal Transformer for Set Supervised Action Segmentation (CVPR2020) https://arxiv.org/abs/2003.14266
38 stars 4 forks source link

BreakFast MoF #2

Open xijing1980 opened 3 years ago

xijing1980 commented 3 years ago

Hi!

Thanks for releasing the code. Also thanks for your well organized code with the comments and documentations. I have tried to replicate the results but I get the MoF of 27% on BreakFast using the I3D features. Is there anything I am missing?

xijing1980 commented 3 years ago

I could get the max MoF of 31.2% on BreakFast and on average 30.3% with 3 trials. I was an older version of the PyTorch and also my I3D extracted features were not correct.

dinggd commented 3 years ago

I could get the max MoF of 31.2% on BreakFast and on average 30.3% with 3 trials. I was an older version of the PyTorch and also my I3D extracted features were not correct.

Can you share your features? Thanks,

ZijiaLewisLu commented 3 years ago

I could get the max MoF of 31.2% on BreakFast and on average 30.3% with 3 trials. I was an older version of the PyTorch and also my I3D extracted features were not correct.

Could you specify how you extracted the feature? Thanks!

I tested on my feature and only obtain very low accuracy.

BIGJUN777 commented 3 years ago

@xijing1980 @MohsenFayyaz89 Could you provide more details about extracting the I3D features? I am really confused about the description "The I3D features are extracted for each frame" mentioned in the paper. Does it work when inputting just a single frame into a 3D CNN model (I3D)?

xijing1980 commented 3 years ago

I tried these features:

https://mega.nz/folder/6nxkFBZb#Fg3Zgy_lIHwUNCnZ6rgj0g

BIGJUN777 commented 3 years ago

@xijing1980 Thank you for your sharing. For those who want to figure out how to extract the features, this issue might be helpful.

AngxiaoYue commented 2 years ago

Hi!

Thanks for releasing the code. Also thanks for your well organized code with the comments and documentations. I have tried to replicate the results but I get the MoF of 27% on BreakFast using the I3D features. Is there anything I am missing?

@xijing1980 Hi! Sorry to disturb you. Did you run the source code directly and did you make any changes? I found that it didn't work very well when I ran the code directly.

AngxiaoYue commented 2 years ago

@ZijiaLewisLu hi! May I ask if you finally raised the results? I ran the source code directly and found that the accuracy was also low?