LibreInstantHMRm-mesh

InstantHMR medium (HGNetV2-B4 backbone, 32.2 M parameters): real-time human mesh recovery in Meta's Momentum Human Rig (MHR), repackaged as a PyTorch checkpoint for LibreYOLO. One person crop in; MHR pose and identity parameters, 70 keypoints in 2D and 3D, the camera translation and the 18,439-vertex mesh out.

⚠️ NON-COMMERCIAL WEIGHTS

These weights are CC BY-NC 4.0, as released by InstantHMR's authors, because part of their training images (SA-1B, AI Challenger) is non-commercial. They are not covered by LibreYOLO's MIT license. Their training labels come from the SAM 3D Body dataset (Meta, SAM License), whose terms are passed through here.

For commercial use, take LibreInstantHMRm-mesh-commercial (Apache-2.0).

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreInstantHMRm-mesh.pt")          # people are found with RF-DETR by default
result = model("photo.jpg", save=True)
meshes = result.meshes                  # vertices, joints3d, joints2d, MHR parameters, camera frame

Fine-tune on your own images with 2D person keypoints (YOLO-pose, COCO-17):

model.train(data="your-pose.yaml", epochs=20)

Export the crop network plus body to ONNX: model.export(format="onnx").

Source

Converted from InstantHMR, https://github.com/Naturalpad-studio/InstantHMR (code Apache-2.0, commit 5437400d4b25c1972821e51916fac34bd96acce6), weights file instanthmr-medium.onnx from momolesang/InstantHMR at revision 07988f51c8c8908cfc7bb44fc18f645c9551b0bf. Authors: Mohamed Adjel, Kahina Chalabi, Vincent Bonnet ("InstantHMR: Real-Time Human Mesh Recovery on Your Phone", HAL preprint hal-05854387).

The checkpoint also carries the MHR body constants built from Meta's MHR rig (Apache-2.0).

Modifications

InstantHMR publishes ONNX only. weights/convert_instanthmr_weights.py in the LibreYOLO repository rebuilds the PyTorch state dict from the graph: folded BatchNorm layers are rebuilt (with running statistics measured on coco128 person crops, eval output unchanged), pre-transposed linear weights are transposed back, and folded constants are unfolded. The converted model matches the released ONNX graphs within 1e-4 on every output.

Accuracy

PA-MPJPE in mm (lower is better), as reported by the InstantHMR authors: EMDB-1 48.8, 3DPW 40.0.

License

CC BY-NC 4.0 (non-commercial) for the weights; SAM License terms for the training labels. See LICENSE and NOTICE.