relsgg-convnext-large

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: RelateAnything: Real-Time Open-Vocabulary Relation Prediction From Any Inputs). Trained on RA-4M; evaluated with OV-SGG-Bench.

Use it

pip install git+https://github.com/Maelic/RelateAnything
hf download maelic/relsgg-convnext-large          # 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-convnext-large", 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-convnext-large", 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), relateanything.onnx, 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	0.527	0.269	0.357
psg	0.387	0.299	0.337
indoorvg	0.542	0.331	0.411
hicodet	0.465	0.310	0.372

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	0.559	0.331	0.416
psg	0.290	0.280	0.285
indoorvg	0.551	0.350	0.428

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

Macro AUC over predicates: 0.7125

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

source	spatial R@50 / mR@50	semantic R@50 / mR@50
vg150	0.630 / 0.308	0.493 / 0.293
psg	0.611 / 0.555	0.424 / 0.317
indoorvg	0.613 / 0.355	0.430 / 0.336

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	0.890	0.330	3592
in front of	0.865	0.335	3576
wearing	0.980	0.678	3416
to the right of	0.830	0.387	3198
to the left of	0.850	0.368	3100
resting on	0.970	0.572	2165
on	0.915	0.453	2040
holding	0.975	0.464	1552
beside	0.980	0.186	1401
next to	0.945	0.236	1352
above	0.885	0.323	1278
below	0.895	0.315	1239
part of	0.905	0.493	1135
supporting	0.985	0.186	933
looking at	0.960	0.262	871

Provenance

run	relsgg-convnext-large
git	e10f70a7256936058af4e3cc6d1f3fe1e69c5aac
backbone	facebook/dinov3-convnext-large-pretrain-lvd1689m
text student	runs/packed/text_student_v2_512/student.pt sha256 e0317830b68ea51e...
ONNX opset / parity	17 / max
torch / transformers	2.14.1+cu130 / 5.19.0
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}
}