Reproduction of the numbers in the paper

[hongsukchoi]

Thanks again for your great work. But I found two critical bugs.

1. While the paper says VIBE uses a bi-directional GRU, the code and pre-trained model weights you provide are set to uni-directional GRU. I think this is a critical difference, and either the code or the paper should be revised. Besides, I can't reproduce the numbers on 3DPW as https://github.com/mkocabas/VIBE/issues/99#issue-650045015. When I train VIBE using bi-GRU with the code and datasets provided here, I get MPJPE: 92.3440, PA-MPJPE: 57.3929, PVE: 108.6150, ACCEL: 25.3047, ACCEL_ERR: 26.3427, which is worse than the baseline.

2. I can't get the same numbers on MPI-INF-3DHP test set with your pre-trained model weights (vibe_model_wo_3dpw.pth.tar). I evaluated 17 joints of valid frames of MPI-INF-3DHP test set aligning the hip joint, and got MPJPE: 101.1546, PA-MPJPE: 65.2166, PCK: 78.74. This is clearly worse than MPJPE: 97.7 PA-MPJPE: 63.4 which is reported in your paper. Also, VIBE's PCK (89.0) reported in the paper is too high compared to SPIN's (76.4). Could you explain VIBE's evaluation process on MPI-INF-3DHP test?

Thank you.

[uyoung-jeong]

I got different result. MPJPE: 99.1523, PA-MPJPE: 64.8514, PVE: 912.4414, ACCEL: 31.4988, ACCEL_ERR: 31.2689. Still not the same as the paper.