What cause the large performance improvement on 3DPW dataset?
linjing7 opened this issue · comments
Hi, thank you very much for your excellent work.
In the journal version, the performance of PyMAF is significantly better than that in the conference version. Could you please tell me the reason? Is it caused by the differences between the training setting?
Hi, thanks for your questions.
Yes, it is caused by the differences between the training settings used in the conference version (following the setting in SPIN) and the journal version (following the setting in PARE). Basically, the performance of the baseline (HMR architecture) is significantly better than the conference version when following the same training setting as PARE. In comparison with SPIN, the training setting of PARE mainly differs in the following aspects:
- Using the EFT pseudo ground-truth labels.
- Using the weights pretrained in pose_resnet for a 2D pose estimation task to initialize the backbone network.
- Using different dataset ratios (i.e. first 100% COCO-EFT then mixture datasets) in the training.
More details about the training of PARE can be found in its paper:
The issues of the numerical evaluation and its fairness are also discussed in our survey paper.
Okay, thank you very much for your reply!