Libre3DMOODt

The official 3D-MOOD Swin-T checkpoint, exposed through LibreYOLO's detect3d API.

Source

Derived from cvg/3D-MOOD at 41bb2904932d91507338e75ae4c802d67616ca1b. Copyright 2025 Computer Vision and Geometry Lab, ETH Zurich. Licensed under Apache-2.0.

The source checkpoint is gdino3d_swin-t_120e_omni3d_699f69.pt from the official model repository at revision 3d1fab552189f1a62fdb60ebb606d20625a30b90, which declares Apache-2.0.

Modifications

The learned parameters and upstream serialization are unchanged. The file is renamed to Libre3DMOODt.pt for LibreYOLO's canonical size convention. SHA-256: 699f69454625d1fb3ad7d3669d8d662d52c349c99b2aba28da1215b8e7e4a555.

3D-MOOD runs through a separately installed upstream runtime. LibreYOLO does not use the undeclared-license vis4d_cuda_ops extension; its adapter selects a permissively licensed portable PyTorch attention path.

Use

import numpy as np
from libreyolo import Libre3DMOOD

model = Libre3DMOOD(
    size="t",
    runtime_path="/path/to/3D-MOOD",
    runtime_python="/path/to/3D-MOOD/.venv/bin/python",
)
result = model("image.jpg", intrinsics=np.load("intrinsics.npy"), text=["chair", "table"])
print(result.boxes3d.xyz)
print(result.depth_map.mean)

The integration is inference-only. Camera intrinsics are required. Accuracy claims remain those of the upstream project; LibreYOLO's completed checks cover runtime parity and output mapping.

License

Apache License 2.0. See LICENSE and NOTICE.