search

kenshohara / 3D-ResNets-PyTorch

3D ResNets for Action Recognition (CVPR 2018)
MIT License
3.9k stars 930 forks source link

i get low accracy on Kinetics #89

Open lee2h opened 6 years ago

lee2h commented 6 years ago

from your paper, you get top-1 accuracy 60.1% on resnet-34, but i get 57.9%. i use your pre-train model. the result is video accuracy. what's the problem?

i can't download some of kinetics dataset video. because some videos are blocked. reason of low accracy is that? or any other reason?

lee2h commented 6 years ago

also, on UCF101, i get the accuracy 82%, 85% on pretrained ResNet18 and Resnet34. how get the your paper accuracy 84%, 87%?.

finetune on layer4 and fc (conv5_x and fc).

victorkkkk commented 6 years ago

How do you process the image? I think your image pre-process method may be different with kenshohara's. I face the same problem, and I get only 20% accuracy when I test kenshohara's UCF pretrain model on UCF dataset directly. (I wrote the pre-process by myself, only use cv2.cvtColor( ) ) @lee2h Could you please explain your image process method? @kenshohara

lee2h commented 6 years ago

@victorkkkk maybe not. i use kenshohara's resnet-18 (pretrained on Kinetics and he do training on UCF101 split 1). i get 84% accuracy, same as paper's result. but, i do training on UCF101 (pretrained on Kinetics from kenshohara), i get 82%.

i do not change any pre-processing code.

cientgu commented 5 years ago

@victorkkkk maybe not. i use kenshohara's resnet-18 (pretrained on Kinetics and he do training on UCF101 split 1). i get 84% accuracy, same as paper's result. but, i do training on UCF101 (pretrained on Kinetics from kenshohara), i get 82%.

i do not change any pre-processing code.

I got the same issue, did you solve it? Many thanks.