laura-wang / video_repres_mas

code for CVPR-2019 paper: Self-supervised Spatio-temporal Representation Learning for Videos by Predicting Motion and Appearance Statistics
MIT License
62 stars 10 forks source link

Train from scratch #3

Open luodezhao opened 5 years ago

luodezhao commented 5 years ago

Hi, thanks for your sharing again! I tried do train C3D from scratch (without pretrained model), I got an accuracy of 81% which is different from what mentioned in your paper. Do you know what caused the difference?

laura-wang commented 5 years ago

what kind of model do you use? And 81% is the clip accuracy? Actually, our number ~45% is from Du tran's C3D paper, and we use the same model as him.

luodezhao commented 5 years ago

Hi, thanks for your reply. The score of 81 is the train accuracy...sorry for that. and could you share your fine-tune code from c3d? Thank you~