LibreZipDepthbnpu-depth

ZipDepth base with the NPU decoder: a ~6M-parameter reparameterizable CNN (RepVGG encoder, SPPF + cross-scale neck, FPN decoder) for zero-shot relative monocular depth estimation, repackaged for LibreYOLO. This checkpoint carries a separately trained unfold-free convex-upsampling head for NPU/edge compilers that lack gather/unfold support; use it when exporting to constrained runtimes. The standard GPU/CPU checkpoint is LibreZipDepthb-depth.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreZipDepthbnpu-depth.pt")  # auto-downloads from this repo
results = model.predict("image.jpg", save=True)
depth = results[0].depth_map.data                # (H, W) relative inverse depth
model.export(format="onnx")                      # unfold-free fixed-resolution graph

Outputs follow LibreYOLO's depth task contract: a dense float map on the original image canvas, higher values mean closer to the camera, no metric unit.

Source

Derived from fabiotosi92/ZipDepth (checkpoints/zipdepth_base_npu.pth) at commit 6b96f4d205f8a2e5377e81c1b74cc99a47f6693a (ECCV 2026, University of Bologna). Copyright (c) 2026 Fabio Tosi. Licensed under the MIT License.

Modifications

LibreYOLO checkpoint-schema metadata wrap only. Learned parameters are byte-identical to upstream. See weights/convert_zipdepth_weights.py in the LibreYOLO source repository.

Training-data provenance

Upstream trained these weights by distilling pseudo-labels produced by Depth Anything V2 Large (a CC-BY-NC-4.0 checkpoint) over ~14M images from 17 public datasets. The MIT grant on the resulting student weights is the upstream authors' published licensing position; the lineage is documented here for transparency.

License

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