I got a 100% accuracy in ucf101 test_list2 and test_list3 by using ucf101_rgb weights

shuangshuangguo / caffe2pytorch-tsn

Transform the caffe model to pytorch model for Temporal Segment Network

73 stars 16 forks source link

I got a 100% accuracy in ucf101 test_list2 and test_list3 by using ucf101_rgb weights #17

Open DUANJINHAO opened 5 years ago

DUANJINHAO commented 5 years ago

Hi, thanks for your great job. I got a comparable result on ucf101-split1 after loading ucf101_rgb.pth weights. Strangely, I got 100% accuracy on split2 and split3 respectively. All the results were computed by using test_models.py provided from tsn-pytorch repository, and I'm sure it works. Finally, I got 95.33% accuracy averaging on these three splits which is totally different from the paper(Only RGB). Could you please give me the results that you test on split2 and split3? Thanks again.

shuangshuangguo commented 5 years ago

The model given in this repo is trained on split1, so you need to test it on split1 too, not split2 and split3. I do not support models on split2 and split3.

DUANJINHAO commented 5 years ago

All right, thaks :)

DUANJINHAO commented 5 years ago

Hi @shuangshuangguo For UCF101 dataset, there is an official classInd.txt file. But I cannot find this file in HMDB51. Could you please tell me where to get that? Thanks

DUANJINHAO commented 5 years ago

Looks like in an alphabetic order.

shuangshuangguo commented 5 years ago

Yes, you are right.