Skip to content

THIRD_PARTY_LICENSES.md

Latest commit

 

History

History
80 lines (69 loc) · 6.01 KB

THIRD_PARTY_LICENSES.md

File metadata and controls

80 lines (69 loc) · 6.01 KB
Raw

Third-Party Licenses

MindSight depends on several third-party libraries and embedded components. This file documents their licenses for compliance purposes.

Embedded Components

MGaze / gaze-estimation

  • License: MIT
  • Copyright: (c) 2024 Yakhyokhuja Valikhujaev
  • Location: mindsight/GazeTracking/Backends/MGaze/gaze-estimation/
  • Full license: mindsight/GazeTracking/Backends/MGaze/gaze-estimation/LICENSE

Gazelle (Gaze-LLE)

  • License: MIT
  • Copyright: (c) 2024 Fiona Ryan
  • Location: Plugins/GazeTracking/Gazelle/gazelle/
  • Full license: Plugins/GazeTracking/Gazelle/gazelle/LICENSE

ptgaze normalization math (vendored)

  • License: MIT
  • Copyright: (c) 2017 hysts
  • Location: mindsight/GazeTracking/normalized/ (canonical MediaPipe 468-point 3D face model data, head-pose fitting, normalizing-warp and gaze-denormalization math from hysts/pytorch_mpiigaze_demo, vendored for the head-pose-normalized estimator backends)

Python Dependencies

Package License Notes
ultralytics AGPL-3.0 YOLO/YOLOE detection engine. The combined work must comply with AGPL-3.0 network-use provisions.
PyQt6 GPL-3.0 GUI framework (GPL-licensed edition)
CLIP (ultralytics fork) MIT Vision-language model, installed via git
torch BSD-3-Clause Deep learning framework
torchvision BSD-3-Clause Vision utilities for PyTorch
onnxruntime MIT ONNX model inference engine
opencv-python Apache-2.0 Computer vision library
opencv-contrib-python Apache-2.0 Extended OpenCV modules
uniface MIT RetinaFace face detector
mediapipe Apache-2.0 MediaPipe inference
numpy BSD-3-Clause Numerical computing
pandas BSD-3-Clause Data manipulation
scikit-learn BSD-3-Clause Machine learning (used by vendored Gazelle utils)
matplotlib PSF/BSD Plotting
Pillow HPND (Historical Permission Notice and Disclaimer) Image handling
PyYAML MIT YAML parsing
requests Apache-2.0 HTTP client
rich MIT Terminal formatting
tqdm MPL-2.0 Progress bars
typer MIT CLI framework
click BSD-3-Clause CLI utilities
networkx BSD-3-Clause Graph algorithms
sympy BSD-3-Clause Symbolic math
huggingface_hub Apache-2.0 HuggingFace model hub integration

External Models (Loaded at Runtime)

Model License Source
DINOv2 Apache-2.0 Meta AI, loaded via torch.hub by the Gazelle backend
MobileGaze weights (*_gaze.onnx / *.pt) MIT (yakhyo/gaze-estimation release) Trained on Gaze360, a non-commercial research dataset -- treat the weights as research use only. The manifest carries this as license_note and the Models tab surfaces it.
Gaze-LLE checkpoints MIT (fkryan/gazelle release) Trained on GazeFollow / VideoAttentionTarget (research datasets); DINOv2 backbone Apache-2.0.
Gaze-LLE DINOv3-distilled ONNX (pico / n / femto / atto / ViT tiny-plus) MIT (PINTO0309/gazelle-dinov3 release) Blend-path gaze-target models (the HGNetV2 pico/n/femto/atto tiers and the DEIMv2 ViT tiny-plus tier). Backbones are Apache-2.0 (D-FINE HGNetV2 / DEIMv2 ViT) trained on DINOv3 outputs via distillation -- no Meta weights embedded. Trained on GazeFollow / VideoAttentionTarget (research datasets). Downloaded from the upstream release; not redistributed by MindSight.
Gaze-LLE DINOv3 ONNX (ViT-S/16) MIT (PINTO0309/gazelle-dinov3 release) + DINOv3 License Embeds the Meta DINOv3 ViT-S/16 backbone weights, so use is additionally subject to the DINOv3 License (Copyright (c) Meta Platforms, Inc.): commercial use is permitted; redistributing the weights or derivatives requires passing through the DINOv3 License text and attribution. MindSight does not redistribute these weights -- the manifest downloads them from the upstream release on demand. Trained on GazeFollow / VideoAttentionTarget (research datasets).
YOLO / YOLOE weights AGPL-3.0 Ultralytics, auto-downloaded on first use
FastSAM-s (Weights/SAM/) AGPL-3.0 CASIA-IVA-Lab FastSAM via the Ultralytics wrapper and asset release; powers the VP Builder's Suggest mode. Same license class as the ultralytics package itself.
RetinaFace weights MIT (uniface release) yakhyo/uniface, auto-downloaded to ~/.uniface on first use; trained on WIDER FACE
MPIIFaceGaze checkpoint (mpiifacegaze_resnet_simple.pth) MIT (hysts/pytorch_mpiigaze_demo release) Trained on MPIIFaceGaze (CC BY-NC-SA, non-commercial research dataset) -- treat the weights as research use only. The manifest carries this as license_note and the Models tab surfaces it. Powers the opt-in --mpiifacegaze-model backend.
MediaPipe Face Landmarker (face_landmarker.task) Apache-2.0 Google MediaPipe model storage (versioned path); 468-point landmarker feeding the head-pose-normalized gaze backends.
Intel gaze-estimation-adas-0002 (gaze-estimation-adas-0002.onnx) Apache-2.0 Intel Open Model Zoo model, converted by MindSight from the published OpenVINO IR to ONNX (numerically verified against the OpenVINO runtime) and redistributed under Apache-2.0 as a MindSight release asset. Powers the opt-in --adas-gaze-model backend.
ArcFace embeddings (--face-reid-sim) MIT (uniface release) yakhyo/uniface, auto-downloaded on first use. Upstream provenance: InsightFace model zoo, which marks its models for non-commercial research use; the training set (WebFace600K) is research-only. Off by default; enable only if that provenance is acceptable for your use.

AGPL-3.0 Notice

MindSight is licensed under AGPL-3.0, consistent with its dependency on ultralytics (AGPL-3.0). If you distribute or provide network access to this software, you must make the complete corresponding source code available under the same license. See the AGPL-3.0 license for full details.