@camstack/addon-ai 0.4.123 → 0.4.124

package/THIRD_PARTY_MODELS.md

@@ -0,0 +1,233 @@
+<!--
+  THIRD_PARTY_MODELS.md: the product notice for every model CamStack can download.
+
+  HAND-WRITTEN FOR NOW. Task 7 of docs/superpowers/plans/2026-09-26-model-license-compliance.md
+  replaces this file with generator output (scripts/gen-model-licenses.ts), built from the
+  licence records on the model catalogs. Until then this root file is the single source:
+  - packages/addon-pipeline, packages/addon-post-analysis, packages/addon-ai and server/backend
+    each carry a byte-identical COPY (npm cannot ship a file from outside the package).
+    Edit this root file, then copy it:  for d in packages/addon-pipeline packages/addon-post-analysis
+    packages/addon-ai server/backend; do cp THIRD_PARTY_MODELS.md "$d/"; done
+    scripts/check-model-notices-shipped.ts fails CI when a copy drifts.
+  - The Hugging Face mirror (camstack/camstack-models) receives this same file, staged by
+    hf/camstack-models/upload.sh. Its model card points at the "Source offer" section below.
+  Facts: docs/superpowers/specs/2026-09-26-model-license-compliance.md (verified 2026-09-26).
+  Not legal advice.
+-->
+
+# Third-party models
+
+CamStack's own source code is MIT-licensed (see `LICENSE`). CamStack does not
+include model weights in its npm packages or its Docker image. It downloads
+them, on demand, from the locations listed below. **Each model stays under its
+own licence.** Converting a model to another format does not change its
+licence: every CamStack-made build is a modified version (a "model derivative")
+of the upstream weights, and the change is stated per model.
+
+Full licence texts: `licenses/models/` in the CamStack source repository, and
+[`LICENSES/`](https://huggingface.co/camstack/camstack-models/tree/main/LICENSES)
+in the model mirror https://huggingface.co/camstack/camstack-models.
+
+Some of these models are **not licensed for commercial use**. They are listed
+under [Non-commercial models](#non-commercial-models).
+
+## Required attributions
+
+- **Ultralytics YOLO26, YOLOv9 and YOLOv8** — © Ultralytics. Licensed under
+  AGPL-3.0 (`AGPL-3.0.txt`). See the [source offer](#source-offer-for-the-agpl-30-weights).
+  https://github.com/ultralytics/ultralytics
+- **YOLOv9** (architecture) — Chien-Yao Wang, I-Hau Yeh, Hong-Yuan Mark Liao,
+  GPL-3.0 (`GPL-3.0.txt`). https://github.com/WongKinYiu/yolov9
+- **SCRFD / InsightFace** — code MIT. The pretrained weights are released by
+  InsightFace for non-commercial research purposes only (`InsightFace-terms.txt`).
+  https://github.com/deepinsight/insightface
+- **MobileCLIP** — © Apple Inc. *Apple Machine Learning Research Model is
+  licensed under the Apple Machine Learning Research Model License Agreement.*
+  (`Apple-AMLR.txt`). CamStack's MobileCLIP files are **Model Derivatives**:
+  the image encoders are converted to OpenVINO IR and CoreML at FP16; the text
+  encoders are quantised to INT8 ONNX and converted to OpenVINO IR.
+  https://github.com/apple/ml-mobileclip
+- **Llama 3.2** — **Built with Llama.** Llama 3.2 is licensed under the Llama
+  3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
+  (`Llama-3.2-Community.txt`; Acceptable Use Policy:
+  https://www.llama.com/llama3_2/use-policy)
+- **fast-plate-ocr** — Copyright (c) 2024 ankandrew, MIT (`MIT-fast-plate-ocr.txt`).
+  https://github.com/ankandrew/fast-plate-ocr
+- **RF-DETR** — © Roboflow, Apache-2.0. https://github.com/roboflow/rf-detr
+- **YuNet** — weights Copyright (c) 2020 Shiqi Yu, MIT (`MIT-yunet-reference.txt`).
+  https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet ·
+  [license](https://github.com/opencv/opencv_zoo/blob/main/models/face_detection_yunet/LICENSE).
+  Training code BSD-3-Clause (`BSD-3-Clause.txt`),
+  https://github.com/ShiqiYu/libfacedetection.train.
+  Trained on WIDER FACE, whose images are CC BY-NC-ND.
+- **AuraFace v1** — fal.ai, Apache-2.0. https://huggingface.co/fal/AuraFace-v1
+- **EasyOCR** — JaidedAI, Apache-2.0. https://github.com/JaidedAI/EasyOCR
+- **U²-Net** — Xuebin Qin et al., Apache-2.0. https://github.com/xuebinqin/U-2-Net
+- **YAMNet** — © Google LLC, Apache-2.0.
+  https://github.com/tensorflow/models/tree/master/research/audioset/yamnet
+- **AIY Vision Birds V1** and the **Coral Edge TPU models** — © Google LLC,
+  Apache-2.0. https://github.com/google-coral/test_data
+- **EfficientNet-Lite0** via **timm** — Ross Wightman, Apache-2.0.
+  https://github.com/huggingface/pytorch-image-models
+- **torchvision MobileNetV3** — BSD-3-Clause (`BSD-3-Clause.txt`). https://github.com/pytorch/vision
+- **Scrypted plugin models** — Koushik Dutta, repository tagged MIT
+  (`MIT-scrypted-plugin-models.txt`). https://huggingface.co/scrypted/plugin-models
+- **Qwen2.5, Qwen3-VL, Qwen3.6** — Alibaba Cloud, Apache-2.0. https://huggingface.co/Qwen
+- **Open Images V7** (animal-classifier training data) — annotations © Google
+  LLC, licensed under CC BY 4.0 (`CC-BY-4.0.txt`). Images are listed by Google
+  as CC BY 2.0, without warranty.
+  https://storage.googleapis.com/openimages/web/factsfigures_v7.html
+- **Vehicle classification dataset** (vehicle-type-v1 training data) — by
+  DrBimmer, https://huggingface.co/datasets/DrBimmer/vehicle-classification,
+  licensed under CC BY-NC 4.0 (`CC-BY-NC-4.0.txt`). CamStack fine-tuned a model
+  on it; the resulting weights are treated as non-commercial.
+- **AudioSet** ontology and labels (YAMNet) — © Google LLC, CC BY 4.0.
+  https://research.google.com/audioset/
+- **COCO** annotations (YOLO pretraining) — COCO Consortium, CC BY 4.0. Images
+  remain under the terms of their Flickr owners. https://cocodataset.org/#termsofuse
+
+## Source offer for the AGPL-3.0 weights
+
+<!--
+  This section is the ONE place the source offer is written. The package copies are
+  byte-identical copies of this file, and the Hugging Face model card, its LICENSE and its folder
+  cards link here instead of repeating it.
+-->
+
+The folders `objectDetection/yolo26`, `objectDetection/yolov9`,
+`segmentation/yolo26-seg`, `plateDetection/yolov8-plate` and
+`packageDetection/yolov8-package` of the model mirror hold Ultralytics weights,
+modified by CamStack and distributed under AGPL-3.0. Their Corresponding Source
+is the public CamStack source repository, **https://github.com/camstack/server**
+(MIT), together with the upstream checkpoints it starts from:
+
+1. **The upstream checkpoints**, published by Ultralytics at
+   https://github.com/ultralytics/assets/releases: `yolo26{n,s,m,l,x}.pt`,
+   `yolov9{t,s,m,c}.pt`, `yolo26{n,s,m}-seg.pt`.
+2. **The export scripts** in https://github.com/camstack/server:
+   `scripts/build-camstack-models.py` (ONNX with a dynamic batch axis, CoreML
+   FP16, OpenVINO IR, OpenVINO NNCF INT8) and `scripts/export-models.py`. The
+   build instructions (Python environment, invocation, calibration images for
+   INT8) are in each script's header.
+3. **The software** that downloads and runs these weights, CamStack itself, is
+   the same repository.
+
+What the repository does **not** cover yet:
+
+- **Some hosted sizes and input resolutions.** The export scripts do not list
+  every variant in the mirror (for example the `-320`/`-256` YOLO26 exports,
+  `yolov9m`/`yolov9c`, and the `yolo26*-seg` models). Each one is the same
+  Ultralytics `export()` with a different model id or `imgsz`, but no script in
+  the repository reproduces it as-is.
+- **`yolov8n-plate` has no training checkpoint.** It is a third-party model
+  (trained 2023-12-05 according to its own metadata, origin not identified).
+  The ONNX file is the only form CamStack has.
+- **The `yolov8n-package` checkpoint is not published yet.** CamStack
+  fine-tuned it with Ultralytics.
+
+## Non-commercial models
+
+These are downloaded by default or selectable, and their terms **do not allow
+commercial use**:
+
+| Model | Terms |
+| --- | --- |
+| `scrfd-2.5g` (face detection, selectable, not the default since 2026-09-26) | InsightFace pretrained models: non-commercial research purposes only |
+| `mobileclip-s1`, `mobileclip-s2`, `mobileclip-s1-text`, `mobileclip-s2-text` (semantic search, `mobileclip-s1` default) | Apple Machine Learning Research Model License: research purposes only, no commercial product or service |
+| `vehicle-type-v1` (vehicle classifier) | trained on a CC BY-NC 4.0 dataset; weights treated as non-commercial |
+| `bird-nabirds-404` (hidden, pending removal) | NABirds: no reproduction, distribution or products without the Cornell Lab's written permission |
+
+## Models
+
+"Mirror" means https://huggingface.co/camstack/camstack-models, folder given.
+"Code / weights" gives the two licences separately because they often differ.
+**Unknown** means no licence could be traced; it does not mean free to use.
+
+### Object detection
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `yolo26n`, `yolo26s`, `yolo26m`, `yolo26l`, `yolo26x`, with every `-fp16`, `-int8`, `-320`, `-256`, `-320-int8`, `-256-int8` variant | Ultralytics YOLO26, COCO-pretrained | AGPL-3.0 / AGPL-3.0 | COCO 2017 (annotations CC BY 4.0) | exported to ONNX, OpenVINO IR (FP32/FP16) and CoreML FP16; `-320`/`-256` re-exported at that input size; `-int8` OpenVINO NNCF post-training INT8 | Mirror `objectDetection/yolo26` |
+| `yolov9{t,s,m,c}-320`, `yolov9{t,s,m,c}-640`, and their `-int8` twins | Ultralytics YOLOv9 weights, COCO-pretrained; architecture by WongKinYiu | AGPL-3.0 / AGPL-3.0 (original YOLOv9 repository GPL-3.0) | COCO 2017 | exported to ONNX, OpenVINO IR FP16 and CoreML at 320 and 640; `-int8` NNCF INT8 | Mirror `objectDetection/yolov9` |
+| `ssd-mobilenet-v2-coco-edgetpu`, `efficientdet-lite0-edgetpu` | Google Coral `test_data` | Apache-2.0 / Apache-2.0 | COCO | none | https://github.com/google-coral/test_data |
+| `ssdlite-mobiledet-coco-edgetpu` (selectable, not the default) | Google Coral `test_data`, `ssdlite_mobiledet_coco_qat_postprocess_edgetpu.tflite` (sha256 `b69e508e…434a2e`); MobileDets (Xiong et al., CVPR 2021) trained with the TensorFlow Object Detection API, https://github.com/tensorflow/models/tree/master/research/object_detection | Apache-2.0 / Apache-2.0 (`Apache-2.0.txt`; https://github.com/google-coral/test_data/blob/master/LICENSE) | COCO 2017 (annotations CC BY 4.0, images under Flickr terms) | none: mirrored byte for byte, with the CPU build and `coco_labels.txt` beside it (`scripts/build-ssdlite-mobiledet-model.py`) | Mirror `objectDetection/ssdlite-mobiledet` |
+
+### Package detection
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `rfdetr-package` | Roboflow RF-DETR-M | Apache-2.0 / Apache-2.0 | fine-tuning set not recorded (unverified) | fine-tuned to 2 classes at 576; exported to ONNX, OpenVINO IR and CoreML; training checkpoint `.pth` also published | Mirror `packageDetection/rfdetr-package` |
+| `yolov8n-package` | Ultralytics YOLOv8n | AGPL-3.0 / AGPL-3.0 | most likely Roboflow Universe `package-at-front-door` (unverified) | fine-tuned by CamStack with Ultralytics; exported to ONNX, OpenVINO IR and CoreML | Mirror `packageDetection/yolov8-package` |
+
+### Face detection
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `scrfd-2.5g` | InsightFace SCRFD 2.5G (bnkps), via the ONNX re-upload `RuteNL/SCRFD-face-detection-ONNX` | MIT / **InsightFace: non-commercial research only** | WIDER FACE (research-oriented terms) | converted to OpenVINO IR and CoreML FP16 | Mirror `faceDetection/scrfd` |
+| `yunet-2023mar` (face-detection default since 2026-09-26) | OpenCV Zoo YuNet 2023mar, `face_detection_yunet_2023mar.onnx` (fp32), https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet | BSD-3-Clause (training code, https://github.com/ShiqiYu/libfacedetection.train; `BSD-3-Clause.txt`) / MIT, Copyright (c) 2020 Shiqi Yu (opencv_zoo; `MIT-yunet-reference.txt`; https://github.com/opencv/opencv_zoo/blob/main/models/face_detection_yunet/LICENSE) | WIDER FACE (images CC BY-NC-ND) | input contract baked into the graph: the pool's RGB /255 becomes the upstream BGR 0..255 (`Mul(255)` plus a channel `Gather`); internal tensor names cleared in the OpenVINO IR; ONNX fp32, OpenVINO IR fp32, CoreML fp16 (`scripts/build-yunet-model.py`) | Mirror `faceDetection/yunet` |
+| `scrypted-yolov9t-face` (retired from the catalog, D643; nothing downloads it) | Scrypted `plugin-models` (YOLOv9t ReLU face) | repository tagged MIT; YOLOv9 architecture GPL-3.0 upstream / **Unknown** per model | unknown | OpenVINO IR converted by CamStack | its OpenVINO IR is still on mirror `faceDetection/scrypted-yolov9-face` |
+| `ssd-mobilenet-v2-face-edgetpu` (hidden) | Google Coral `test_data` | Apache-2.0 / Apache-2.0 | — | none | https://github.com/google-coral/test_data |
+
+### Face recognition
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `arcface-r100` (a ResNet34) | re-upload `onnx-community/arcface-onnx` (itself pointing at `garavv/arcface-onnx`); original author not stated | **Unknown** / **Unknown** | not stated | converted to OpenVINO IR and CoreML FP16 | Mirror `faceRecognition/arcface` |
+| `auraface-r100` | fal.ai AuraFace v1 (`glintr100.onnx` only) | Apache-2.0 / Apache-2.0 | stated upstream as commercially and publicly available sources | input scaling `y = 2x - 1` baked ahead of the first Conv (`scripts/build-auraface-model.py`); converted to OpenVINO IR and CoreML FP16 | Mirror `faceRecognition/auraface` |
+| `inception-resnet-v1` | Scrypted `plugin-models` | repository tagged MIT / **Unknown** per model | not stated | OpenVINO IR converted by CamStack | ONNX and CoreML: https://huggingface.co/scrypted/plugin-models · OpenVINO: mirror `faceRecognition/inception-resnet-v1` |
+
+### Licence plate detection and OCR
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `yolov8n-plate` | Ultralytics YOLOv8n fine-tuned for plates by a third party (trained 2023-12-05 per its metadata) | AGPL-3.0 / AGPL-3.0 | **unknown** | converted to OpenVINO IR, CoreML and TFLite fp32 (`tflite/camstack-yolov8n-plate_float32.tflite`, mirrored but not referenced by the catalog) | Mirror `plateDetection/yolov8-plate` |
+| `cct-s-v2-global` | `ankandrew/fast-plate-ocr` release model | MIT / MIT | private, not released upstream | input head removed (recorded in the ONNX `doc_string`); converted to OpenVINO IR and CoreML | Mirror `plateRecognition/cct-s-v2-global` |
+| `vgg-english-g2` | JaidedAI EasyOCR, via Scrypted `plugin-models` | Apache-2.0 / Apache-2.0 | — | OpenVINO IR converted by CamStack | ONNX and CoreML: https://huggingface.co/scrypted/plugin-models · OpenVINO: mirror `plateRecognition/vgg_english_g2` |
+
+### Classifiers
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `animal-classifier` | timm EfficientNet-Lite0 | Apache-2.0 / Apache-2.0 | Open Images V7 animal subset (annotations CC BY 4.0, © Google LLC) | fine-tuned by CamStack (8 classes); exported to ONNX, OpenVINO IR and CoreML | Mirror `animalClassification/animal-classifier` |
+| `animals-10` (hidden, pending removal) | origin not identified | **Unknown** / **Unknown** | unknown | converted to ONNX, OpenVINO IR and CoreML | Mirror `animalClassification/animals-10` |
+| `bird-classifier` | Google AIY Vision `birds_V1` | Apache-2.0 / Apache-2.0 | — | converted from TensorFlow with tf2onnx; OpenVINO IR and CoreML derived from it | Mirror `animalClassification/bird-classifier` |
+| `bird-classifier-aiy-edgetpu` (hidden) | Google Coral `test_data` | Apache-2.0 / Apache-2.0 | — | none | https://github.com/google-coral/test_data |
+| `bird-nabirds-404` (hidden, pending removal) | ResNet50 trained on NABirds (origin not recorded) | **Unknown** / NABirds: no products without Cornell Lab permission | NABirds (Cornell Lab of Ornithology) | converted to ONNX, OpenVINO IR and CoreML | Mirror `animalClassification/bird-nabirds` |
+| `vehicle-type-v1` | torchvision MobileNetV3-Large | BSD-3-Clause / **CC BY-NC 4.0** (from its data) | DrBimmer `vehicle-classification`, CC BY-NC 4.0 | fine-tuned by CamStack (8 classes); exported to ONNX, OpenVINO IR and CoreML | Mirror `vehicleClassification/vehicle-type-v1` |
+| `vehicle-type-efficientnet` (hidden, pending removal) | trained on VMMRdb (origin not recorded) | **Unknown** / VMMRdb terms unverified | VMMRdb | converted to ONNX, OpenVINO IR and CoreML | Mirror `vehicleClassification/efficientnet` |
+
+### Semantic search (CLIP)
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `mobileclip-s1`, `mobileclip-s2` (image) | Apple `ml-mobileclip` | MIT / **Apple Machine Learning Research Model License** | per Apple | Model Derivative: converted to OpenVINO IR and CoreML FP16 | Mirror `clip/mobileclip-s1`, `clip/mobileclip-s2` |
+| `mobileclip-s1-text`, `mobileclip-s2-text` | Apple `ml-mobileclip`, with the CLIP tokenizer | MIT / **Apple Machine Learning Research Model License** | per Apple | Model Derivative: quantised to INT8 ONNX and converted to OpenVINO IR | Mirror `clip/mobileclip-s1`, `clip/mobileclip-s2` |
+
+### Segmentation
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `u2netp` | `xuebinqin/U-2-Net` | Apache-2.0 / Apache-2.0 | DUTS-TR (terms not checked) | converted to ONNX, OpenVINO IR and CoreML | Mirror `segmentationRefiner/u2netp` |
+| `yolo26n-seg`, `yolo26s-seg`, `yolo26m-seg` | Ultralytics YOLO26-seg, COCO-pretrained | AGPL-3.0 / AGPL-3.0 | COCO 2017 | exported to ONNX, OpenVINO IR and CoreML | Mirror `segmentation/yolo26-seg` |
+
+### Audio
+
+| Catalog ids | Upstream | Code / weights | Training data | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- | --- |
+| `yamnet-onnx` | `tensorflow/models` `research/audioset/yamnet` | Apache-2.0 / Apache-2.0 | AudioSet (labels CC BY 4.0, © Google LLC) | converted from TensorFlow with tf2onnx; OpenVINO IR derived from it | Mirror `audioClassification/yamnet` |
+| `apple-soundanalysis` | Apple SoundAnalysis, built into macOS | under Apple's macOS licence | — | nothing downloaded | built in |
+
+### Language and vision-language models (AI addon, not hosted by CamStack)
+
+| Catalog ids | Upstream | Code / weights | CamStack changes | Downloaded from |
+| --- | --- | --- | --- | --- |
+| `llm-qwen2.5-1.5b-instruct-q4` | Qwen2.5 1.5B Instruct | — / Apache-2.0 | none (GGUF Q4_K_M by Qwen) | https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF |
+| `llm-llama3.2-3b-instruct-q4` | Meta Llama 3.2 3B Instruct | — / Llama 3.2 Community License | none by CamStack (GGUF Q4_K_M by bartowski) | https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF |
+| `llm-qwen3-vl-2b-instruct-q4`, `llm-qwen3-vl-4b-instruct-q4`, `llm-qwen3-vl-8b-instruct-q4` | Qwen3-VL Instruct | — / Apache-2.0 | none by CamStack (GGUF by unsloth) | https://huggingface.co/unsloth (`Qwen3-VL-{2B,4B,8B}-Instruct-GGUF`) |
+| `llm-qwen3.6-35b-a3b-ud-q4` | Qwen3.6 35B-A3B | — / Apache-2.0 | none by CamStack (GGUF UD-Q4_K_M by unsloth) | https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF |
+
+### Your own models
+
+A model you import through Model Studio (uploaded, converted, or pulled from
+Frigate+ or Scrypted) keeps whatever licence its author published. CamStack
+does not change it and does not record it.