Seeking Guidance on ZJU-Mocap Dataset Split and Hyperparameter Settings

[jingi0614]

Dear Authors,

First of all, thank you very much for sharing your excellent work and for maintaining this repository. I have been greatly inspired by your paper, and I am currently trying to reproduce the reported results by implementing a custom data loader and following the configurations described in the paper and code.

However, despite my best efforts, the performance of my reproduction falls short of the numbers presented in the paper. I would like to kindly ask for your guidance on two main points, especially regarding the ZJU-Mocap dataset:

1. Exact dataset split (ZJU-Mocap)
   • Could you please clarify the train/validation/test split for ZJU-Mocap that was used in the paper?
   • For example:
   • Which subjects (e.g., CoreView_xxx IDs) were assigned to train/val/test?
   • How many frames and which views per subject were included?
   • Were there any filtering rules (excluded frames, invalid cameras, or occlusion thresholds)?
   • Having access to the exact split files (JSON/CSV) would be extremely helpful for faithful reproduction

2. Hyperparameters or training details not explicitly mentioned
   • I closely followed the paper and repository defaults (optimizer, LR schedule, loss weights, augmentation), but there may be important hyperparameters or dataset-specific overrides that are not explicitly documented.
   • Could you kindly confirm:
   • The final set of loss weights (photometric, perceptual, regularizers, geometry terms).
   • Any dataset-specific adjustments for ZJU-Mocap (e.g., intrinsic scaling after cropping, focal recalibration, near/far plane handling).

3. My setup and results (for reference)
   • Environment: <Ubuntu 22.04 / CUDA 11.8 / PyTorch 2.0.1>
   • GPU: RTX 4090 (24GB)
   My reproduction results on ZJU-Mocap are as follows: {"psnr": 30.1178, "ssim": 0.9613, "lpips": 0.0459}

Although these numbers are quite reasonable, they are still different from the results reported in the paper. I kindly ask if there might be any additional hyperparameters, training details, or specific dataset splits (especially for ZJU-Mocap) that could explain this gap.

Thank you very much again for your time and kind support. I truly appreciate your guidance.

With best regards,

[mu-you302]

The ZJU-Mocap dataset I downloaded from neuralbody only has six subjects(my_377 my_386 my_387 my_392 my_393 my_394), while this paper states:

For ZJU-MoCap, we select 7 subjects for training and the other 3 for testing.

I'm not sure where the problem is.
How many subjects did you use in your implementation?