LibreGTRs-depth

GTR-S monocular depth weights converted for LibreYOLO.

GTR support is being prepared for LibreYOLO v1.6.0. Earlier PyPI releases may not include this model family.

Usage

With a LibreYOLO version that includes GTR depth:

from libreyolo import LibreYOLO
model = LibreYOLO("LibreGTRs-depth.pt")
result = model.predict("image.jpg")[0]
inverse_depth = result.depth_map

The default input is a 640 by 640 stretch resize. This checkpoint supports the depth task only.

Output

Results.depth_map is relative inverse depth on the original image canvas: higher values are closer. The network's native output is metric depth from a log-depth head (exp(clamp(logit, -4, 5)) metres); LibreYOLO returns its exact reciprocal, so 1 / depth_map recovers the metre estimate. That scale holds only for cameras and scenes like the training data; treat it as relative otherwise.

Source

Official GTR implementation, source revision 782e737efe2e6437ac537fbdcee089673d3376c1.

Published checkpoint, weight repository revision 9fc62c8c2b2c976835d0f1c1ffc544dbc0f9e29f. Copyright (c) 2026 Intellindust-AI-Lab. The source code is MIT licensed and the publisher's weight repository explicitly declares MIT.

The model was pretrained by its authors on a mixed metric-depth corpus (SUN RGB-D, DIODE, Virtual KITTI 2, KITTI, Hypersim, TartanAir, ARKitScenes and ImageNet pseudo-labels).

Modifications

Selected the EMA state dict and added LibreYOLO schema v1.0 metadata. Learned parameters and state-dict keys are unchanged. Training/optimizer state was removed. Conversion uses weights/convert_gtr_weights.py in the LibreYOLO source repository.

Validation

Strict loading and exact CPU parity with the pinned upstream graph (maximum absolute difference 0.0 on the metre output at 640px, portable attention) were checked for this artifact, along with CPU prediction and fixed-resolution ONNX/TorchScript export. No independently reproduced benchmark accuracy or latency numbers are claimed here.

License

MIT. See LICENSE and NOTICE.