Closed W4ngH4o closed 2 years ago
Can you provide more details about the model and dataset? If you test the model on Kinetics, please check this issue to map the label.
Thanks for your reply. I did have a different kinetics_400_categroies.txt (the contents in it are sorted according to A-Z),maybe that's why the result is close to 0. However, the link of your kinetics_400_categroies.txt is 404, please help me access it. Thanks so much.
Thanks. I remap the label and test the model. However, the top1-acc is 64.22. I wonder if the input method of data affect the result. In fact, I extracted frames from videos firstly, and then stored in SSD. Then, it is similar to sampling Something dataset. It's just not random sampling in a segment like TSN, but sampling every 8 frames, if it's 8x8.
If you change the dataset.py
, please adopt the same dense sampling strategies as used in pre-trained model. The same frame and same stride.
Thanks so much. I use dataset.py by you to decode videos, and the top1_acc is 74. But I want to know why we need to convert fps to the same fps before frame sampling. And why the target video fps is selected as 30. Thanks for your reply.
The target video fps 30 is the default fps for Kinetics videos. We use fps30 for a fair comparison.
As there is no more activity, I am closing the issue, don't hesitate to reopen it if necessary.
I have downloaded the pretrain model for testing UniFormer-S. Why the top_acc is close to 0?