search

kenshohara / 3D-ResNets-PyTorch

3D ResNets for Action Recognition (CVPR 2018)
MIT License
3.89k stars 931 forks source link

Low accuracy on UCF Sports Action dataset. #85

Open ArchieGu opened 6 years ago

ArchieGu commented 6 years ago

Hi, I am using the UCF Sports Action dataset which is kind of a subset of UCF101. And I followed the instruction in Readme file to prepare the dataset. However after I trained 200 epoch, the accuracy of both train and test are pretty low (30%train and 22% test)

Could you give me some advices about the data preparing or the possible reason about the low accuracy?

kenshohara commented 6 years ago

Did you perform finetuning using our pretrained models? What did you use training configurations?

ArchieGu commented 6 years ago

No, I didn't use the finetuning models yet. I will try it asap.

By the way, what is the accuracy you achieved on UCF101 dataset by using none finetuning model on resnet 34?

I can only achieve accuracy around 38% on validation, and the training accuracy can achieve around 80%. I think it goes overfitting. Do you have any suggestion about this?

Thank you .

kenshohara commented 6 years ago

When I train ResNet-18 on UCF-101 from scratch, I got around 80% and 40% accuracies on the training and validation sets, respectively. Because UCF-101 is a relatively small dataset to train 3D CNNs from scratch, it causes overfitting. I think the training on UCF-Sports is similar.

ArchieGu commented 6 years ago

Oh, right, UCF-Sports is really a very small dataset. I think I can train the model using UCF-101 and using UCF-Sports to test the model.

ArchieGu commented 6 years ago

One more question, why I cannot using resnet 101 to train the model? Each time I set depth !=34, the program gives me an error about illegal access to memory.

86