shapes-mhr

MHR (Momentum Human Rig) articulated body model fits exported by the src/sam3d_grounding/run.py pipeline in sam3d-for-shape-corr.

Directory Layout

shape-mhr/
└── <dataset>/               # one subfolder per dataset (e.g., faust_r)
    ├── mhr/                 # MHR outputs — one pair of files per input mesh
    │   ├── <name>.obj           # Aligned MHR mesh
    │   └── <name>_mhr_params.npy  # MHR parameters & alignment transforms
    └── off/                 # Original input meshes (copied from source dataset)
        └── <name>.off

File Formats

mhr/<name>.obj — Aligned MHR Mesh

A triangulated Wavefront OBJ mesh of the fitted MHR body model, aligned to the corresponding input shape.

Property	Value
Vertices	18 439 (fixed — shared topology across all shapes)
Faces	36 874 (fixed — shared topology across all shapes)
Coordinate unit	metres (same scale as the source dataset)
Axis convention	Y-up, same as source dataset axis order

Because all shapes in a dataset share the same MHR topology, vertex index i of mesh A corresponds to the same anatomical location as vertex index i of mesh B — enabling direct vertex-to-vertex correspondence across shapes.

mhr/<name>_mhr_params.npy — Parameters

A NumPy .npy file saved with allow_pickle=True containing a Python dict with the following keys:

Key	Shape	dtype	Description
identity_coeffs	(B, 45)	float32	MHR shape identity coefficients (body, head, hands)
model_parameters	(B, 204)	float32	MHR joint angles and scalings
face_expr_coeffs	(B, 72)	float32	MHR facial expression blendshape coefficients
R	(B, 3, 3)	float32	Per-view rotation matrices (MHR → input mesh space)
t	(B, 3)	float32	Per-view translation vectors
s	(B,)	float32	Per-view scale scalars

B is the number of camera views used in the final surface optimisation stage (default: 3 best views selected by Chamfer distance).

Important — params vs. mesh: The exported .obj is produced by generating a mesh from each of the B parameter sets independently, then averaging their vertex coordinates (not the parameters). The B rows in identity_coeffs / model_parameters are therefore B distinct parameter solutions — each a plausible fit optimised from a different camera viewpoint. When using these params for training or inference, treat each row as an independent candidate solution rather than as a batch that should be averaged or pooled together.

Loading example:

import numpy as np

params = np.load("tr_reg_000_mhr_params.npy", allow_pickle=True).item()
identity_coeffs  = params["identity_coeffs"]   # (B, 45)
model_parameters = params["model_parameters"]  # (B, 204)
R, t, s          = params["R"], params["t"], params["s"]

Generation Pipeline

Each MHR mesh is produced by a 6-step pipeline:

1. Multi-view rendering — input mesh rendered from multiple camera angles (azimuth × elevation grid).
2. SAM 3D Body inference — keypoints and depth extracted per view.
3. Keypoint triangulation — 3D keypoints triangulated from multi-view 2D detections.
4. Keypoint optimisation — MHR model_parameters optimised to match triangulated keypoints.
5. Surface optimisation — identity_coeffs and model_parameters jointly refined using Chamfer distance against the (optionally trimmed) input mesh, with optional bone-length and keypoint regularisation terms.
6. Multi-view merging — per-view MHR vertices averaged (weighted by Chamfer distance) to produce the final mesh.

Datasets

Subfolder	Source	# Shapes	Input format
faust_r	FAUST remeshed	100	.off

Correspondence

Because MHR meshes share a fixed topology (18 439 vertices, 36 874 faces), a vertex-to-vertex functional correspondence between any two shapes in the same dataset is established directly by index — vertex i in one MHR mesh maps to vertex i in any other.

This can be exploited for:

• Spectral (functional map) correspondence methods
• Texture transfer across shapes
• Training data generation for shape correspondence networks