nrs-large / ATTRIBUTION.md
Pyassahu's picture
nrs-large 0.1.0 (Apache-2.0): weights, model card, licences
309a787 verified
|
Raw History Blame Contribute Delete
2.15 kB

Attribution: nrs-large 0.1.0

This file credits every work nrs-large 0.1.0 is built from. The exact licence texts are in LICENSES/; the short legal notice is NOTICE.

Licence of the nrs-large 0.1.0 weights and scripts: Apache License, Version 2.0 (LICENSE).

1. Code and network design

The network was trained by Necessity Robotics from random initialisation. Only the authors' MIT-licensed code design is used; no DLNR weights were used for training.

Work Used for Licence Copyright Please cite
DLNR / High-frequency-Stereo-Matching-Network, github.com/David-Zhao-1997/High-frequency-Stereo-Matching-Network, commit c10475f nrs-large network MIT (LICENSES/DLNR_MIT.txt) (c) 2021 Princeton Vision & Learning Lab (the RAFT-Stereo licence the repository carries) Zhao et al., High-Frequency Stereo Matching Network, CVPR 2023; Lipson, Teed, Deng, RAFT-Stereo, 3DV 2021

2. Training data

The only training data is WMGStereo-150k. Its licence statement is a tag on the dataset card (see the table).

Source Licence Evidence Credit and citation Used
WMGStereo-150k (indoor, nature_filtered, flying), Hugging Face revision 5ecc0dcd BSD-3-Clause bsd-3-clause tag on the Hugging Face dataset card (present since the first author upload); the generator repository carries the BSD-3-Clause text (LICENSES/WMGStereo_generator_BSD-3-Clause.txt). If the dataset licence changes, we will stop distributing these weights and retrain without it. Yan, Raistrick, Deng, What Makes Good Synthetic Training Data for Zero-Shot Stereo Matching?, CVPR 2026, arXiv 2504.16930. Notice: Copyright (c) 2025, Princeton Vision & Learning Lab, with the three conditions and the disclaimer yes, the only training data (155,769 pairs)

Not training data:

  • Held-out monitor only: 200 TartanAir V1 pairs from the hold-out environments (CC BY 4.0; Wang et al., TartanAir, IROS 2020; CMU AirLab). They scored the run's progress and selected nothing.
  • Evaluation only, never trained on: Middlebury, ETH3D, KITTI 2012/2015, Booster.