jhmdb acc?

[zgs731]

How do you train, so that the network accuracy rate is above 77%, and the network I train is only about 70%. The data set is 21 categories, not 14 categories. I hope you can explain the training method.

[fandulu]

Plz clone this code and run, I forget to fix a bug on colab, sorry for it.

Update: I fixed the bug on colab now, could you try it again? Besides, anhminh3105 uses this code to pulls val_acc of each split over 80% on JHMDB with the weighted class to alleviate the skewed effect since 'walk' has ~3 times more data compared to others.

[zgs731]

Thank you for your work！@fandulu I got the accuracy close to you at colab. I want to try anhminh3105's work,but I didn't find his code. and 'walk' has only 41 segments,basically consistent with other classes.

[fandulu]

@zgs731 , yes, I checked as the sample number is as follows, anhminh3105 may use this model on his own data and I misunderstood it, sorry for the misleading comments. brush_hair:41 catch:48 clap:44 climb_stairs:40 golf:42 jump:39 kick_ball:36 pick:40 pour:55 pullup:55 push:42 run:40 shoot_ball:40 shoot_bow:53 shoot_gun:55 sit:39 stand:36 swing_baseball:54 throw:46 walk:41 wave:42

[anhminh3105]

Hi @fandulu and @zgs731.

It was my mistake to say that class 'walk' has 3 times more data in JHMDB as it was the case only for the HMDB dataset of which the JHMDB is a subset and I was looking over the HMDB homepage and was referencing it over based on this image.

I just also uploaded my variation to github and if you would like to check it out, please find it here. Just to justify how I got val_acc above 80%, it was because I only trained the model on a subset of JHMDB dataset with added weighted class training. Specifically, the dataset subset contained only specified classes (e.g. walk, sit, catch, etc.) that I could easily test it using Openpose using input from camera with myself.

Br.

[fandulu]

Hi, anhminh3105, thanks very much for offering these comments!