# Third-party notices

The **code** in this repository is licensed under AGPL-3.0-only (see `LICENSE`
and the additional terms in `NOTICE`).
The **released model weights** and the **RA-4M annotations** are derivatives of
third-party assets and carry the additional terms below. Nothing in this file is
legal advice; each upstream license is authoritative.

## Model weights

| Component | Upstream | Terms that travel with our artifacts |
|---|---|---|
| Vision backbones (ViT-S/16, ViT-S/16+, ViT-B/16; ConvNeXt in ablations) | Meta **DINOv3** pretrained weights (`facebook/dinov3-*`, gated) | Released RelSGG checkpoints are **derivatives of DINOv3 weights** and are distributed under the [DINOv3 License](https://ai.meta.com/resources/models-and-libraries/dinov3-license/), not AGPL-3.0. Each model card states this. |
| Predicate text encoder (distilled student) | Distilled from Meta **dino.txt** (DINOv3 text tower) outputs; tokenizer is OpenAI **CLIP** BPE (MIT) | The student checkpoint is a distillation derivative of dino.txt; DINOv3 license terms apply. |

## Training data

| Component | Source | Terms |
|---|---|---|
| **RA-4M annotations** (all relation labels in the `maelic/RA-4M` dataset) | Generated by **google/gemma-4-26B-A4B-it** | Generated outputs are distributed with the notice required by the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). Downstream users must comply with the Gemma Prohibited Use Policy when using these annotations. |
| MegaSG **images** | Objects365 / COCO / OpenImages re-crawl | **Images are NOT redistributed.** We publish image identifiers/URLs only; each image remains under its source dataset's license (Objects365 research terms, COCO CC BY 4.0, OpenImages CC BY 2.0/4.0 per image). |
| `vg_raw` relation annotations | **Visual Genome** | CC BY 4.0 — attribution: Krishna et al., *Visual Genome*, IJCV 2017. |
| Eval packs (VG150, PSG, IndoorVG, HICO-DET, SpatialSense) | Respective benchmark releases | Used for evaluation; redistributed annotations retain their upstream licenses and citations (see paper bibliography). |

## Reel photographs

`assets/reel/images/` holds the six photographs `deploy/make_reel.py` renders
into `assets/reel/hero.gif`. All are from **Wikimedia Commons** and are
redistributed here under their own terms; per-file title, author, licence and
source page are in `assets/reel/credits.json`, which is the authoritative list.

The README's video reel is a separate artifact, rendered by
`deploy/render_video.py` from five Commons **video** clips. Those clips are NOT
redistributed here — `deploy/fetch_footage.py --from_shots` downloads them from
Commons — and their titles, authors, licences and source pages are in
`assets/reel_video/credits.json`. One of the five is CC BY-SA, so the rendered
reel is an adaptation and carries CC BY-SA 4.0.

| Clip | Author | Licence |
|---|---|---|
| R3C Adoption (54729571599) | blmcalifornia | Public domain |
| Gof a Gweithiwr Metrel Ray Burrell | Rupertmaesglas | CC0 |
| Technical Hardtail riding | Radde faehrt Radd | CC BY 3.0 |
| Jalebi being prepared, Bangalore | Subhashish Panigrahi | CC BY-SA 3.0 |
| David Farmer guitarist at Copped Hall | Acabashi | CC BY-SA 4.0 |

| Photograph | Author | Licence |
|---|---|---|
| `bicycle.jpg` | Clem Onojeghuo | CC0 |
| `catlaptop.jpg` | acebrand | CC0 |
| `frisbee.jpg` | Sally Wehner | CC BY 2.0 |
| `skateboard.jpg` | Infrogmation of New Orleans | CC BY-SA 4.0 |
| `horse.jpg` | Subhrajyoti07 | CC BY-SA 4.0 |
| `tennis.jpg` | Chris Eason | CC BY 2.0 |

The last two also ship with the browser demo as sample photos. The CC BY and
CC BY-SA images require attribution wherever the rendered video appears; the
project page carries it in its footer.

## Detectors (demo)

The webcam demo uses YOLO-World v2 / YOLOE / YOLO26 detectors built on **ultralytics**
(**AGPL-3.0**). Detector weights are **not** part of the RelSGG release artifacts
(the AGPL-3.0 code and the DINOv3-licensed relation checkpoints). They are shipped
separately, and only there:

* The public **browser demo** (github.com/Maelic/RelateAnything_demo, served at
  maelic.github.io/RelateAnything_demo) hosts ONNX exports of `yoloe-11s-seg(-pf)`,
  `yolov8s-worldv2` (re-parameterized to RA-4M-497) and `yolo26n/s` as separate
  model files under **AGPL-3.0**, listed in that repo's `THIRD_PARTY.md`. The page
  loads them at runtime; the corresponding source is ultralytics at the version
  named there plus `deploy/web/export_web_models.py` here.
* Everything else (laptop bundles, HF release) still rebuilds locally:

      deploy/reparam_detector.py        # re-parameterize to the RA-4M-497 vocabulary
      deploy/export_detector_onnx.py    # export the reparameterized detector to ONNX

Users of either path are bound by AGPL-3.0 for the detector component. The RelSGG
relation model does not link against ultralytics at inference time (numpy +
onnxruntime / onnxruntime-web only). The MobileCLIP-BLT text tower used to
re-parameterize YOLOE is likewise only distributed with the browser demo, under
Apple's MobileCLIP weights license.

## Not redistributed, referenced only

- Meta DINOv3 raw pretrained checkpoints (gated download from Meta).
- `mobileclip_blt.ts` (Apple MobileCLIP, fetched separately).
- SAM / SAM3 segmentation checkpoints used in data generation.