search

kenshohara / 3D-ResNets-PyTorch

3D ResNets for Action Recognition (CVPR 2018)
MIT License
3.84k stars 930 forks source link

Low accuracy on Kinetics (2D resnet is better than 3D resnet) #88

Open gurkirt opened 5 years ago

gurkirt commented 5 years ago

I trained a 2D resnet101 on kinetics dataset, and it has much better accuracy than your 3D resent. I used https://github.com/gurkirt/2D-kinectics to train 2D resnet. Of course, pretrained on imagenet.

I got the top1 accuracy of 70.3 on kinetics-400.
Do you think it is because of pretraining? Have you tried 2D networks with your training procedure?

kenshohara commented 5 years ago

Thank you for your interesting information. I have not tried training 2D networks on my own. I think the pretraining on ImageNet influences the difference but it was huge gap... I also want to try your code and consider the difference.

bityangke commented 5 years ago

In the Non-local Neural Network paper, the top 1 acc of 2D-ResNet50 baseline is 71.8. @gurkirt @kenshohara I believe the performance gap does not come from the ImageNet pretraining as @kenshohara said.

kkjh0723 commented 5 years ago

@kenshohara @gurkirt @bityangke Did anyone find the reason for the difference?