augmentation in feeder about 2s-agcn HOT 7 CLOSED

I have tried the augmentation but found no improvement for NTU-RGBD. For Skeleton-Kinetics it is important.
The influence of the input length is not investigated.
I haven't tried the 120 dataset.

from 2s-agcn.

I tried to change the input length to 64. For joint stream, I got 94.59 top1 acc. And for bone stream, I got 94.11 top1 acc. The final fusion result is 95.53.

from 2s-agcn.

I tried to change the input length to 64. For joint stream, I got 94.59 top1 acc. And for bone stream, I got 94.11 top1 acc. The final fusion result is 95.53.

Divide the input sequence into 64 equal segments, and then choose a random frame from each segment?@VSunN

from 2s-agcn.

@Lixmdeep , I just changed the window size in the configuration to 64.

from 2s-agcn.

@VSunN I can not reproduce your result, can you tell me any other configurations you have changed in the codefile? thanks alot.

from 2s-agcn.

@hongge831 set random_choose=True and window_size=64 in config/nturgbd-cross-view/train_joint.yaml

from 2s-agcn.

@hongge831 set random_choose=True and window_size=64 in config/nturgbd-cross-view/train_joint.yaml

The paper only mentioned to set T to 300 without any other augmentation, whether using random choose be a kind of augmentation? Can I set it to be True to reproduce the results? Thx!

from 2s-agcn.