Unable to produce same quantitative results that are shown in paper

[aakashjuseja-aj]

I have trained in two setting

1. first stage pre-trained and second stage from scratch I get, MPJPE: 69.09 : P-MPJPE: 54.66
2. Training both stages from scratch results in even worst. (I am using 60 epochs for first stage, how many epochs would you suggest?)

Can you please help me how to get the results that are shown in paper. I am also using 1024 batchsize and 100 epoch but unable to replicate the results!

[FanJunqiao]

(1) Currently the dataloader has some bug that may concatenate frames of different sequence together, which leads to slight performance drop. I am trying to find and fix the bug.
(2) Training first stage should try to avoid overfitting, can try first train a few epoch (1 epoch for mmBody) and use the feature for second stage training.

[aakashjuseja-aj]

I have used fewer epochs for stage as well. Results are still worst. I have given my best, I believe results are not reproducible.