search

mzolfaghari / ECO-efficient-video-understanding

Code and models of paper " ECO: Efficient Convolutional Network for Online Video Understanding", ECCV 2018
MIT License
437 stars 96 forks source link

validation precision of pytorch-ECO-Lite on Something-something V1 #39

Closed ntuyt closed 5 years ago

ntuyt commented 5 years ago

Hi, using the pytorch code provided. I only obtain 38.3 accuracy on Something-something V1 validation set. Which is worse than reported, any suggestions?

ntuyt commented 5 years ago

By the way, I use 16 segments settings.

pencilzhang commented 5 years ago

By the way, I use 16 segments settings.

I am also using the pytorch code with 16 segments and obtain 39.3% top-1 accuracy on sth-sth v1 val set. Did you figure it out?

ntuyt commented 5 years ago

No, I have not figured out. It is unsolved.

mzolfaghari commented 5 years ago

@ntuyt @pencilzhang Unfortunately PyTorch version does not give exact accuracy we obtained by Caffe. We were able to reproduce similar performance on other datasets but for the Sth-sth_v1 the result is lower. This is is due to some minor difference between the two implementations we found later.

Please don't hesitate to reopen the issue if you had further questions.