relsgg-vitb16

Open-vocabulary relation prediction from any boxes or masks. Give the model an image and regions from any source (a detector, a segmenter, ground truth); it returns ranked relations over a predicate vocabulary supplied at inference, and optionally two graphs (spatial + semantic) from the same forward pass. Object class labels are never an input.

Part of RelateAnything (code · paper). Trained on RA-4M; evaluated with OV-SGG-Bench.

Use it

pip install git+https://github.com/Maelic/RelateAnything
hf download maelic/relsgg-vitb16          # optional; the API fetches on first use

from relsgg import RelateAnything

# Regions come from any detector, any segmenter, or your own annotation.
# Object class labels are never an input.
model = RelateAnything.from_pretrained("maelic/relsgg-vitb16", device="cuda")
for t in model.predict(image, boxes_xyxy, topk=20):    # PIL/ndarray, boxes [N, 4] in pixels
    print(t)                                           # (person) --riding [0.67]--> (horse)

# Masks instead of boxes: pass the [N, H, W] binary masks beside their extents.
triplets = model.predict(image, boxes_xyxy, masks=masks, topk=20)

# The vocabulary is an input. Any strings, at any time, without retraining.
model.set_vocabulary(["about to collide with", "reflected in"])

# Or answer from the whole training vocabulary, 19,103 strings, read from the weights.
model = RelateAnything.from_pretrained("maelic/relsgg-vitb16", full_vocabulary=True, device="cuda")

# Two graphs from one forward pass.
graphs = model.predict(image, boxes_xyxy, decompose=True)   # {"spatial": [...], "semantic": [...]}

Every vocabulary is encoded once by the text student shipped beside the weights, and the head is reparameterized onto it; scoring afterwards is vision only. full_vocabulary=True reads predicate_embeddings.npz instead of encoding, which turns a minute and a half of CPU work into a download. model.pth embeds the backbone configuration, so running these weights needs no gated DINOv3 login.

Files: model.pth (torch, EMA weights), text_student.pt + tokenizer, predicate_embeddings.npz (the training vocabulary, encoded), predicate_bank.npz, thresholds.json, calibration.json, README.md.

Every number below is generated from measured eval artifacts (release/make_model_cards.py); none is hand-typed.

Closed-vocabulary transfer (reparameterized, TEST, graph-constrained)

source	R@50	mR@50	F1@50
vg150	0.531	0.289	0.375
psg	0.412	0.302	0.349
indoorvg	0.538	0.302	0.387
hicodet	0.468	0.313	0.375

Open-vocabulary, NO reparameterization (all 19,103 predicates deployed)

Synonym-matched at the calibrated tau (see provenance). This is the honest "the model never saw your label set" protocol.

source	SoftR@50	SoftmR@50	SoftF1@50
vg150	0.566	0.345	0.429
psg	0.324	0.299	0.311
indoorvg	0.545	0.357	0.432

Spatial reasoning (SpatialSense, adversarial true/false; chance = 0.5)

Macro AUC over predicates: 0.6860

Two-graph decomposition (spatial / semantic, type-stratified protocol)

source	spatial R@50 / mR@50	semantic R@50 / mR@50
vg150	0.636 / 0.314	0.494 / 0.312
psg	0.608 / 0.541	0.420 / 0.327
indoorvg	0.617 / 0.359	0.429 / 0.312

Deployment thresholds (per-predicate best-F1, measured on THIS checkpoint)

Score scales are checkpoint-specific (the output head is rank-trained), so these thresholds transfer to no other model. Regime: gt boxes, pair_weight=0, 5000 val images. Top predicates by support:

predicate	threshold	best F1	GT support
behind	0.895	0.337	3598
in front of	0.860	0.343	3580
wearing	0.985	0.684	3417
to the right of	0.860	0.385	3196
to the left of	0.870	0.370	3102
resting on	0.975	0.582	2166
on	0.925	0.460	2043
holding	0.980	0.469	1552
beside	0.980	0.194	1402
next to	0.935	0.231	1352
above	0.895	0.332	1283
below	0.895	0.328	1241
part of	0.905	0.495	1135
supporting	0.985	0.194	945
looking at	0.965	0.271	872

Provenance

run	relsgg-vitb16
git	e9ea42aed60f766f12ad19d51709129c50110a3b
backbone	facebook/dinov3-vitb16-pretrain-lvd1689m
text student	runs/packed/text_student_v2_512/student.pt sha256 e0317830b68ea51e...
ONNX opset / parity	17 / max
torch / transformers	2.13.0+cu130 / 5.14.1
training mixture	megasg_clean + vg_raw + hicodet, per-image 0.727/0.063/0.210; source-aware negatives: ['hicodet']

License and data notices

Weights are a derivative of Meta DINOv3 pretrained weights and are distributed under the DINOv3 license. Training annotations (RA-4M) were generated by gemma-4-26B and carry the Gemma Terms of Use notice; images are referenced by identifier only (Objects365/COCO/OpenImages). The vg_raw subset derives from Visual Genome (CC BY 4.0). Predicate synonyms are deliberately never collapsed — surface-form diversity is part of the label space. Full notices: THIRD_PARTY_NOTICES.md in the code repository.

Citation

@article{neau2026relateanything,
  title   = {RelateAnything: Real-Time Open-Vocabulary Relation Prediction From Any Inputs},
  author  = {Neau, Ma\"elic},
  journal = {arXiv preprint arXiv:2609.12552},
  eprint  = {2609.12552},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url     = {https://arxiv.org/abs/2609.12552},
  year    = {2026}
}