Hi everyone, thanks for releasing such an impressive project. Your work on accelerating the SAM 3D Body pipeline for real-time humanoid robot teleoperation is truly inspiring.
I am currently working on a similar research project involving full-body human mesh recovery and humanoid control (deploying on Unitree G1/similar platforms). I've been studying your paper "Fast SAM 3D Body" and I am particularly interested in the Neural Kinematic Projection module ($f_{\omega}$) described in Section 3.2.
However, after carefully reviewing the repository, I noticed a discrepancy between the paper and the provided code:
- Missing Module: The paper states that the cross-topology translation (MHR to SMPL) is replaced by a lightweight feedforward MLP (approx. 2.5M parameters). However, the current codebase in sam_3d_body/ and tools/ only provides the Conversion class, which appears to utilize iterative PyTorchSMPLFitting (as seen in conversion.py), rather than the feedforward MLP described in the paper.
- Missing Weights: I have also cross-checked the checkpoints/ directory structure mentioned in the README, but I cannot locate the pre-trained weights for this specific MLP projection module.
Could you please clarify if the implementation and pre-trained weights for the Neural Kinematic Projection MLP are intended to be released?If there are specific constraints (e.g., SMPL licensing or internal dependencies) preventing the public release of these specific weights/code, I would greatly appreciate it if you could provide some guidance or hints on the training strategy for the MLP. I am more than happy to re-train the module myself if you could provide the recommended training data configuration beyond what is described in the paper.
Thank you for your time and for this significant contribution to the community.
Hi everyone, thanks for releasing such an impressive project. Your work on accelerating the SAM 3D Body pipeline for real-time humanoid robot teleoperation is truly inspiring.$f_{\omega}$ ) described in Section 3.2.
I am currently working on a similar research project involving full-body human mesh recovery and humanoid control (deploying on Unitree G1/similar platforms). I've been studying your paper "Fast SAM 3D Body" and I am particularly interested in the Neural Kinematic Projection module (
However, after carefully reviewing the repository, I noticed a discrepancy between the paper and the provided code:
Could you please clarify if the implementation and pre-trained weights for the Neural Kinematic Projection MLP are intended to be released?If there are specific constraints (e.g., SMPL licensing or internal dependencies) preventing the public release of these specific weights/code, I would greatly appreciate it if you could provide some guidance or hints on the training strategy for the MLP. I am more than happy to re-train the module myself if you could provide the recommended training data configuration beyond what is described in the paper.
Thank you for your time and for this significant contribution to the community.