Skip to content

ATTRIBUTION.md

Latest commit

 

History

History
78 lines (59 loc) · 3.05 KB

ATTRIBUTION.md

File metadata and controls

78 lines (59 loc) · 3.05 KB
Raw

Attribution and third-party components

License

This repository is distributed under AGPL-3.0-only. It contains derivative YOLOv5/DeformCAT source, and both current upstream repositories identify their code as AGPL-3.0. The exact historical imported revisions are not recoverable, so this release uses the conservative AGPL license instead of claiming a more permissive relicensing right. The root LICENSE controls repository code except where a file or asset states a separate license.

DeformCAT

  • Repository: https://github.com/jiongger/DeformCAT
  • License: AGPL-3.0
  • Role: dual-stream multispectral detector, deformable cross-modal interaction, model graph, training/evaluation scaffold, and KAIST evaluator integration.
  • Locally modified areas include src/models/common.py, src/models/yolo_test.py, src/train.py, src/test.py, and transformer model configs.

The exact imported upstream commit is not recoverable from the original local workspace metadata. This is recorded as a provenance limitation rather than inventing a revision.

YOLOv5

  • Repository: https://github.com/ultralytics/yolov5
  • Current upstream license: AGPL-3.0
  • Role: YOLO-style backbone/head components, training utilities, data loading, loss, plotting, checkpoint, and NMS infrastructure.

The released files retain their upstream-style comments. Original project contributions are enumerated in ARCHITECTURE.md and are not described as authorship of YOLOv5.

COCO API

The full copyright notice is preserved in THIRD_PARTY_NOTICES.md.

KAIST benchmark

  • Repository: https://github.com/SoonminHwang/rgbt-ped-detection

  • Dataset license identified by the official repository: CC BY-NC-SA 4.0

  • Toolbox code license identified by the official repository: BSD-2-Clause

  • Citation:

    Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. “Multispectral Pedestrian Detection: Benchmark Dataset and Baselines.” CVPR, 2015.

The repository does not redistribute the full dataset. The evaluator annotation JSON retained for metric compatibility and the selected qualitative montage containing adapted KAIST frames are separately marked CC BY-NC-SA 4.0 in THIRD_PARTY_NOTICES.md and assets/README.md.

Project contributions

The project owner's work includes:

  • three-scale lightweight DARP-Net integration;
  • consensus/detail and detail-scene reliability modules;
  • ignore-aware and protocol-semantic supervision;
  • KAIST-specific bounded calibration experiments;
  • expert routing and post-processing experiments;
  • cross-dataset diagnostics and fine-tuning;
  • experiment cards, canonical result schema, model cards, release validation, and evidence-grounded LLM reporting.

This attribution file is a best-effort software provenance record, not a substitute for legal advice.