LibreMoGe3l-depth

MoGe-3 ViT-L/14 for monocular metric geometry (depth, point map, camera intrinsics, validity mask), repackaged in LibreYOLO checkpoint format.

Source

Derived from Ruicheng/moge-3-vitl model.pt at revision 184008f877d7ad1ad4c2cd2182a9bd1f63d0e5be, the exact commit LibreYOLO pins. The source model card declares these weights MIT. Copyright (c) Microsoft Corporation. Code: microsoft/MoGe at 74fbce054ebed49800de42d0ad0e83495065719a, MIT License.

The DINOv2 encoder MoGe-3 builds on is separately licensed Apache-2.0 by Meta AI; both notices are preserved in MoGe's composite LICENSE, copied verbatim here.

Modifications

None to the learned parameters: every tensor of the official checkpoint is kept unchanged, and LibreYOLO checkpoint v1.0 metadata (task depth) is added. The same tensors serve the depth and normal tasks. See weights/convert_moge3_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreMoGe3l-depth.pt")
result = model.predict("image.jpg")
depth = result.depth_map            # metric depth in metres, inf where invalid
points = depth.point_map            # (H, W, 3) camera-frame points, metres
K = depth.intrinsics                # (3, 3) OpenCV pixel intrinsics
valid = depth.valid_mask            # (H, W) boolean

License

MIT License. See the LICENSE and NOTICE files in this repository.