THIRD_PARTY_NOTICES.md

Third-party notices

TriDrive builds on the components below. None of their weights or data are redistributed in this bucket unless stated; each keeps its own licence, which applies in addition to ours.

Component	How TriDrive uses it	Licence	Source
V-JEPA 2 ViT-g 384 (facebook/vjepa2-vitg-fpc64-384)	Frozen road-video encoder (road tokens, pooled features)	Code MIT (portions Apache-2.0); weights released by Meta without a separate licence file	https://github.com/facebookresearch/vjepa2
OpenBADAS / BADAS-Open (nexar-ai/BADAS-Open, V-JEPA2-L backbone)	Teacher risk signal for the distilled road student; three warning-probe features	Apache-2.0	https://huggingface.co/nexar-ai/BADAS-Open
Distilled road student checkpoints/deployed/onnx/student_v5/	ViT-S/16 student trained by us on BATON road video to reproduce the pooled V-JEPA 2 feature and the OpenBADAS risk logit	Our weights: CC BY 4.0. As a distillation of the two models above, downstream use must also respect their terms (MIT / Apache-2.0 attribution)	this bucket
MotionBERT (DSTformer code; MB_release weights)	Baseline architecture re-instantiated in code/aide/models/; released-weight fine-tunes are not part of this bucket	Apache-2.0 (attribution and NOTICE requirements apply to the re-instantiated code)	https://github.com/Walter0807/MotionBERT
Driver-WM	Independent re-implementation of the published pipeline in code/aide/ for the head-to-head comparison; our reproduction checkpoints checkpoints/aide/driverwm_repro_s* are trained from scratch on AIDE	Paper: arXiv 2605.05092; no third-party code or weights included	—
YOLO11x-pose (Ultralytics)	Driver detection in the on-vehicle pose chain (called through the ultralytics package; no weights included)	AGPL-3.0	https://github.com/ultralytics/ultralytics
RTMW-x-l (MMPose)	Whole-body pose estimation (frozen; no weights included)	Apache-2.0	https://github.com/open-mmlab/mmpose
openpilot (comma.ai)	Vehicle platform, CAN decoding, and the stock driver-monitoring comparator	MIT	https://github.com/commaai/openpilot
ST-GCN, SiMLPe, VideoMAE v2, L2CS-Net, OpenFace 3.0, MediaPipe	Baseline and feature extractors referenced by the code (no weights included)	See each repository (MIT / Apache-2.0 / OpenFace 3.0 academic licence)	see code/baselines_dms/ registry notes
BATON corpus	Training/evaluation data; the derived features and labels in data/ and benchmark/ are derivatives	CC BY-NC 4.0 (the derivatives here inherit it)	BATON release
AIDE dataset	Keypoint-forecasting benchmark (Table 1); no AIDE data or derived caches are redistributed	Obtain from the AIDE authors	https://github.com/ydk122024/AIDE

Attribution for our own material: "TriDrive authors (anonymous for double-blind review)"; the camera-ready version will replace this line.