Issues with Single-image overfitting

[Xiaoming-Zhao]

Hi @shunsukesaito, thanks a lot for this great work. I am following your guideline to process the free rp_dennis_posed_004 data and tried to overfit to one image.

The only change I made to your repo is that I replaced the following line with self.yaw_list = [0] to always train on the same image: https://github.com/shunsukesaito/PIFu/blob/f0a9c99ef887e1eb360e865a87aa5f166231980e/lib/data/TrainDataset.py#L81

I then trained with the command line your suggested:

python -m apps.train_shape --dataroot {path_to_training_data} --random_flip --random_scale --random_trans

However, after 100 epochs, the following is what I get. May I know whether you have tried single-imge overfitting before? Do you mind sharing some thoughts about why this may fail? Thanks a lot