Wishlist

[caizhongang]

Welcome notes

Please leave a comment on which new method/feature/dataset you would like us to add in MMHuman3D.
The most upvoted suggestions will be prioritized!

Updates

• Requests gathered after the release of v0.3.0.
• Requests gathered after the release of v0.4.0.
• Requests gathered after the release of v0.5.0.
• Requests gathered after the release of v0.6.0.
• Requests gathered after the release of v0.7.0 and v0.8.0.

Links

Please refer to our Roadmap for upcoming updates.

[caizhongang]

In the first three weeks after the release of v0.3.0, these requests have stood out (ordered by popularity):

• Add new SOTAs: PARE and GLAMR (update: PARE is added in v0.7.0)
• Support SOTAs for SMPL-X (Frankmocap, PIXIE, Pose2Pose)
• Support for clothed humans
• Support .fbx output

[lucasjinreal]

small suggestions, if fbx can drive mixamo charactors would be better.

[caizhongang]

@jinfagang Thank you for your suggestion, you may want to start a new comment below instead of commenting on my comment so people can upvote your idea!

[korotetskiiba]

fbx output ❤️

[rlleshi]

When could fbx output be supported? Is there a rough timeline?

[lucasjinreal]

Does it consider support some acceleration method in MMDeploy, for instance, add a more powerful but more faster detector base? Or accelerated pose model from mmpose?

[caizhongang]

@jinfagang Thank you for you suggestion! We are looking into collaboration with other codebases such as MMDeploy and MMRazor. However, we do not have a timeline yet. We will keep you updated.

[socketing]

很希望可以支持rgbd的支持。可以从深度数据上或者点云数据上实现对模型的参数化生成。可以更好的对已生成的三维模型进行尺寸估算

---

Translation: hope to see support for rgbd. Parametric model from depth / point cloud data give better dimension estimation.

[caizhongang]

@socketing 感谢你的建议，基于点云的方法已经在我们的Roadmap中了，相信今年上半年就能加入代码库。

---

Translation: we are developing point cloud-based methods, that are likely to be added in 2022 H1. Please check our Roadmap for updates.

[lucasjinreal]

支持一下HybrIK推理

---

Translation: support HybrIK inference

[rlleshi]

New Feature: Save the predictions (bboxes, poses, meshes) in pkl format

[littlejiumi]

Great work!
Hope to support ROMP: [ICCV 2021] Monocular, One-stage, Regression of Multiple 3D People

[Gzzgz]

Hope to support MANO

[caizhongang]

After the release of v0.4.0, the team has taken notes of the following requests:

• Support point cloud-based methods, suggested by @socketing
• More SMPL SOTA: ROMP, suggested by @littlejiumi
• Save inference results, suggested by @rlleshi (update: added in v0.7.0)
• Collaboration with other codebases (MMRazor, MMDeploy), suggested by @jinfagang

[lucasjinreal]

does there any examples/tutors on deploy any mesh3d human models?

[caizhongang]

• Save inference results, suggested by @rlleshi

@jinfagang we will add a feature to save SMPL inference results, from which you can obtain human mesh vertices and surfaces. You may also refer to this tutorial for visualization of SMPL mesh.

[JackieXuu]

Hope to train SPIN(or other methods) using our own datasets(like openpose-based gt). For SPIN method which needs fitting files, I hope that if we use other own datasets, it can generate initial fitting parameters.

[rawalkhirodkar]

Thank you for the work.
Can we have temporal fitting methods like DynaBOA?

[Kayce001]

Hope to support mobile deployment.

[rlleshi]

In other words, fbx output. Then you can convert that as required, for example to dae and deploy it on an IPhone

[imabackstabber]

Please Add MANO.

[caizhongang]

MANO supported with ExPose in v0.9.0. ;)

[caizhongang]

After the release of v0.5.0, the team has taken note of the following requests:

• Add tools to generate intial fitting parameters for SPIN-like methods, suggested by @ZhengtianXu
• Parametric model beyond body: MANO @imabackstabber @Gzzgz

The following requests have already been taken note in previous request collections:

• Mobile deployment, suggested by @Kayce001
• fbx output, mentioned by @rlleshi

[ykk648]

PyMAF
Relative_Human

[caizhongang]

After the release of v0.6.0, the team has taken note of the following requests:

• More SMPL methods (PyMAF and Relative_Human), suggested by @ykk648

We would like to seek community's kind contribution to MMHuman3D by reproducing SOTA methods! Please start a discussion.

[akk-123]

wish support new sota PIXIE

[JiahongWu1995]

Please Add SmoothNet.

[caizhongang]

SmoothNet has been added in v0.8.0 :)

[Serdnad]

+1 for temporal methods, DynaBOA in particular.

Would also be great to see support for STAR, meant to be "a drop in replacement for SMPL" partially made by one of the original SMPL researchers.

[caizhongang]

STAR is added in v0.9.0. ;)

[ykk648]

shapy

[kristijanbartol]

STRAPS and HierProb3D. These are one of the only methods for accurate shape estimation in-the-wild (most of the other methods often just use an average shape or estimate shape only based on 2D joint estimations).

[caizhongang]

Great suggestion, may also look at Shapy.

[caizhongang]

After the release of v0.7.0 and v0.8.0, the team has taken note of the following requests:

• Temporal fitting methods such as DynaBOA, suggested by @Serdnad
• Shape estimation methods such as STRAPS, HierProb3D, and Shapy, suggested by @ykk648 and @kristijanbartol

Community's contributions to MMHuman3D by reproducing SOTA methods are highly welcome! Please start a discussion.

[ChennyDeng]

Cant wait to see DynaBOA!

[axhiao]

Hi, do you consider this regression based method https://github.com/sail-sg/mvp? Thanks!

[woo1]

Great work!
Hope to support TCMR: [CVPR 2021] Beyond Static Features for Temporally Consistent 3D Human Pose and Shape from a Video
and Hand4Whole: [CVPRW 2022] Accurate 3D Hand Pose Estimation for Whole-Body 3D Human Mesh Estimation

[daixiangzi]

Hope to support cliff

[imabackstabber]

Hope to support MeshTransformer and FastMETRO.
Always overfitting when I train them from scratch.

[ShreelekhaR]

And to support HUMBI or Panoptic Studio as datasets

[Taylorminer]

Hope to support VirtualPose. VirtualPose: Learning Generalizable 3D Human Pose Models from Virtual Data (ECCV 2022)

[YaoBeiji]

啥时候来个4d assocaition graph 多视角多人的3d 估计

[imabackstabber]

Hope to support BEHAVE dataset, which is on cvpr 2023 workshop Rhobin challenge now.

[fpq199182]

请问有人脸3d重建相关的吗