train_refinenet_mt.py

[maicao2018]

1. Is psudo_labels not updated during train_refinenet_mt training?
2. If train_refinenet_mt training works, should the break on line 190 be annotated?

[chaneyddtt]

Hi @maicoa,

1. Yes, the pseudo labels are generated only once at the beginning.
2. Yes, you can either comment line 190 or run the command again without '--generate_pseudol' to train the network.

[maicao2018]

@chaneyddtt Thank you for the above answer. I have another question: does the project only crop the original images, but does not do the same operation on the original keypoints?

[chaneyddtt]

Sorry, I did not get you. Are you referring to the preprocessing?

[maicao2018]

1. Does the final performance of the algorithm largely depend on the accuracy of psudo_labels?
2. According to the protocol you provided, train and test the network, why is its performance lower than CC-SSL?

Mu, Jiteng, et al. "Learning from synthetic animals." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.

[chaneyddtt]

Hi @maicoa,

1. yes, the accuracy of the pseudo labels is important for the final performance.
2. May I know how you train the network and the performance you get?

[chaneyddtt]

Sure, please email to lic@comp.nus.edu.sg.

[maicao2018]

Because the file is more than 25M, it is uploaded through Google Disk. https://drive.google.com/file/d/1vn_uVe23mpt1rqdiub03UNwmp_OThCUt/view?usp=sharing

[chaneyddtt]

Thanks for sharing. I am not very sure about the performance since you are working on a different dataset. I find that the dataset you are using is much smaller than the TigDog, this might lead to the performance drop. And you might also need to tune the hyperparamters, for example the weights for different losses, on this dataset. I also notice from your training log that the performance gradually drops during training, which is a bit wired. Maybe you can check whether the generated pseudo labels are appropriately used during training.