Miayan/stereo-matching-weights

Stereo Matching Model Weights (mirror)

Mirror of three stereo-matching model checkpoints, re-hosted on Hugging Face so they can be pulled from behind a corporate firewall that blocks Google Drive. All weights are copied verbatim from their original public sources — credit and licensing belong to the original authors.

Contents

foundationstereo/ — Fast-FoundationStereo

Source: Google Drive folder 1HuTt7UIp7gQsMiDvJwVuWmKpvFzIIMap. Each timestamped folder is one complete model (model_best_bp2_serialize.pth + cfg.yaml):

Folder	Notes
23-36-37	Most accurate (PyTorch 49.4 ms / TRT 23.4 ms)
20-26-39	Middle
20-30-48	Fastest (PyTorch 38.4 ms / TRT 16.6 ms)
15-44-51	Extra model included in the source folder

onnx/ holds pre-exported ONNX models for 20_26_39, 20_30_48, 23_36_37 at 320x736 / 576x960 and 4 / 8 iterations.

liteanystereo/ — LiteAnyStereo

Source: Google Drive folder 1UvDx296pVk7pC2rozKIpQF_EXcOleZOB.

• LAS1/ — LAS1 checkpoints, incl. LiteAnyStereo.pth (not available on HF elsewhere).
• LAS2/ — LAS2 S/M/H/L (same as tomtomtommi/LiteAnyStereoV2, included for completeness).

mobilestereonet/ — MobileStereoNet

Source: individual Google Drive file IDs (see MobileStereoNet repo). 2D and 3D variants, trained on different combinations of SceneFlow (SF), DrivingStereo (DS) and KITTI2015. *DS+KITTI2015* and *SF+DS+KITTI2015* are the best on KITTI2015 val.

Download example

# NOTE: on the corporate network, add these two lines BEFORE importing
# huggingface_hub, or the TLS interception makes SSL verification fail:
import truststore
truststore.inject_into_ssl()

from huggingface_hub import snapshot_download, hf_hub_download

# whole repo
snapshot_download(repo_id="Miayan/stereo-matching-weights", local_dir="./weights")

# single file
hf_hub_download(repo_id="Miayan/stereo-matching-weights",
                filename="foundationstereo/23-36-37/model_best_bp2_serialize.pth",
                local_dir="./weights")