search

qijiezhao / pseudo-3d-pytorch

pytorch version of pseudo-3d-residual-networks(P-3D), pretrained model is supported
MIT License
451 stars 113 forks source link

About finetuning on UCF101 #7

Closed xxbbml closed 6 years ago

xxbbml commented 6 years ago

Hi~ Thanks for sharing the codes and some experiments results. Recently I have also tried finetuing the P3D199 Kinetics models on UCF101, but only got about 85% accuracy of action recognition task. I just wonder do you have any other tricks or did I make some mistakes..? Thanks!

qijiezhao commented 6 years ago

Yes, training tricks are deeply required if you want to get a high score. I could share you some processes like jittering, TSN and corner-crop testing, every step is necessary.

xxbbml commented 6 years ago

Thanks. I'll try some data augmentation methods. And could you share a little more about how to use TSN on this model? Thanks!

qijiezhao commented 6 years ago

Mainly modify the data input method so that you could select multiple segments together to train for each sample. More details you may take a look at xiong's official TSN repo: https://github.com/yjxiong/tsn-pytorch/blob/master/dataset.py. His code also supports 10-crop testing. FYI

icoz69 commented 6 years ago

hi, did you try any tricks to improve the results? i got similar accuracy