toqi/camtrap-measure-weights

CamTrap Measure weights

Fetched by the CamTrap Measure desktop app at startup via manifest.json. Once the repo is public the app reads it with no token. The app does not download sam3/.

Each component keeps its own licence. The repo as a whole has no single licence, so the card says other.

Commercial use of the distance model (unified/) is not permitted. It is a modified Depth Anything V2 Large model, and Depth Anything V2 Large is licensed CC-BY-NC-4.0 (non-commercial use only).

Folder	Component	Licence	Source and credit
unified/	Distance net (unified distance + CQR)	CC-BY-NC-4.0, non-commercial use only	Modified: fine-tuned from depth-anything/Depth-Anything-V2-Metric-Outdoor-Large-hf (Depth Anything V2, Yang et al., NeurIPS 2024; Base, Large and Giant are CC-BY-NC-4.0 upstream, https://github.com/DepthAnything/Depth-Anything-V2)
speciesnet/	SpeciesNet v4.0.3a (PyTorch)	Apache-2.0	Google; Kaggle google/speciesnet; cite Gadot et al. 2024, IET Computer Vision
roma/dinov2_vitl14_pretrain.pth	DINOv2 ViT-L/14	Apache-2.0	Meta; facebookresearch/dinov2
megadetector/md_v1000.0.0-redwood.pt	MegaDetector v1000 redwood	MIT	agentmorris/megadetector
roma/roma_outdoor.pth	RoMa outdoor	Code MIT (Edstedt et al., CVPR 2024); upstream states no separate licence for the weights	Parskatt/RoMa
sam3/	SAM 3	SAM License (Meta); a copy is in sam3/LICENSE	Meta; facebook/sam3

• unified/ — the paper's unified distance+CQR net (Depth-Anything-V2-Large backbone, 7-channel input, [q05, q50, q95] output); ckpt_unified_scratch_split2_rollfix/best of the research repo.
• speciesnet/ — the always-crop classifier. Its info.json points the detector entry at the MegaDetector file, so nothing else is downloaded.
• roma/ — RoMa outdoor and its DINOv2 ViT-L/14 backbone.
• sam3/ — SAM 3 in transformers format. The app does not download it.

Citations

• Depth Anything V2: L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, H. Zhao. "Depth Anything V2." NeurIPS 2024.
• SpeciesNet: T. Gadot et al. "To crop or not to crop: Comparing whole-image and cropped classification on a large dataset of camera trap images." IET Computer Vision, 2024.
• RoMa: J. Edstedt, Q. Sun, G. Bökman, M. Wadenbäck, M. Felsberg. "RoMa: Robust Dense Feature Matching." CVPR 2024.