README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - scene-graph-generation
  - object-detection
  - visual-relationship-detection
  - pytorch
  - yolo
pipeline_tag: object-detection
library_name: sgg-benchmark
model-index:
  - name: REACT++ yolo12m
    results:
      - task:
          type: object-detection
          name: Scene Graph Detection
        dataset:
          name: VG150
          type: vg150
        metrics:
          - type: mR@20
            value: 10.81
            name: mR@20
          - type: R@20
            value: 18.76
            name: R@20
          - type: mR@50
            value: 14.42
            name: mR@50
          - type: R@50
            value: 24.63
            name: R@50
          - type: mR@100
            value: 16.78
            name: mR@100
          - type: R@100
            value: 28.47
            name: R@100
          - type: F1@20
            value: 13.72
            name: F1@20
          - type: F1@50
            value: 18.19
            name: F1@50
          - type: F1@100
            value: 21.11
            name: F1@100
          - type: e2e_latency_ms
            value: 20.5
            name: e2e_latency_ms
  - name: REACT++ yolo26m
    results:
      - task:
          type: object-detection
          name: Scene Graph Detection
        dataset:
          name: VG150
          type: vg150
        metrics:
          - type: mR@20
            value: 10.81
            name: mR@20
          - type: R@20
            value: 21.12
            name: R@20
          - type: mR@50
            value: 14.6
            name: mR@50
          - type: R@50
            value: 28.34
            name: R@50
          - type: mR@100
            value: 18.36
            name: mR@100
          - type: R@100
            value: 33.7
            name: R@100
          - type: F1@20
            value: 14.3
            name: F1@20
          - type: F1@50
            value: 19.27
            name: F1@50
          - type: F1@100
            value: 23.77
            name: F1@100
          - type: e2e_latency_ms
            value: 19.8
            name: e2e_latency_ms
  - name: REACT++ yolov8m
    results:
      - task:
          type: object-detection
          name: Scene Graph Detection
        dataset:
          name: VG150
          type: vg150
        metrics:
          - type: mR@20
            value: 12.22
            name: mR@20
          - type: R@20
            value: 22.89
            name: R@20
          - type: mR@50
            value: 16.31
            name: mR@50
          - type: R@50
            value: 29.96
            name: R@50
          - type: mR@100
            value: 18.45
            name: mR@100
          - type: R@100
            value: 34.09
            name: R@100
          - type: F1@20
            value: 15.93
            name: F1@20
          - type: F1@50
            value: 21.12
            name: F1@50
          - type: F1@100
            value: 23.94
            name: F1@100
          - type: e2e_latency_ms
            value: 18.7
            name: e2e_latency_ms

REACT++ Scene Graph Generation — VG150 (yolo12m, yolo26m, yolov8m)

This repository contains REACT++ model checkpoints for scene graph generation (SGG) on the VG150 benchmark, across 3 backbone sizes.

REACT++ is a parameter-efficient, attention-augmented relation predictor built on top of a YOLO backbone. It uses:

• DAMP (Detection-Anchored Multi-Scale Pooling), a new simple pooling algorithm for one-stage object detectors such as YOLO
• SwiGLU gated MLP for all feed-forward blocks (½ the params of ReLU-MLP at equal capacity)
• Visual x Semantic cross-attention — visual tokens attend to GloVe prototype embeddings
• Geometry RoPE — box-position encoded as a rotary frequency bias on the Q matrix
• Prototype Momentum Buffer — per-class EMA prototype bank
• P5 Scene Context — AIFI-enhanced P5 tokens provide global context via cross-attention

The models were trained with the SGG-Benchmark framework and described in the REACT++ paper (Neau et al., 2026).

---

Results — SGDet on VG150 test split (CUDA, max_det=100, batch_size=1)

Metrics from end-to-end evaluation (tools/evaluate.py). Latency = model forward only.

Backbone	R@20	R@50	R@100	mR@20	mR@50	mR@100	F1@20	F1@50	F1@100	Lat. (ms)
yolo12m	18.76	24.63	28.47	10.81	14.42	16.78	13.72	18.19	21.11	20.5
yolo26m	21.12	28.34	33.7	10.81	14.6	18.36	14.3	19.27	23.77	19.8
yolov8m	22.89	29.96	34.09	12.22	16.31	18.45	15.93	21.12	23.94	18.7

---

Checkpoints

Variant	Sub-folder	Checkpoint files
yolo12m	yolo12m/	yolo12m/model.onnx (ONNX) · yolo12m/best_model_epoch_19.pth (PyTorch)
yolo26m	yolo26m/	yolo26m/model.onnx (ONNX) · yolo26m/best_model_epoch_18.pth (PyTorch)
yolov8m	yolov8m/	yolov8m/model.onnx (ONNX) · yolov8m/best_model_epoch_6.pth (PyTorch)

---

Usage

ONNX (recommended — no Python dependencies beyond onnxruntime)

from huggingface_hub import hf_hub_download

onnx_path = hf_hub_download(
    repo_id="maelic/REACTPlusPlus_VG150",
    filename="yolo12m/react_pp_yolo12m.onnx",
    repo_type="model",
)
# Run with tools/eval_onnx_psg.py or load directly via onnxruntime

PyTorch

# 1. Clone the repository
#    git clone https://github.com/Maelic/SGG-Benchmark

# 2. Install dependencies
#    pip install -e .

# 3. Download checkpoint + config
from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="maelic/REACTPlusPlus_VG150",
    filename="yolo12m/best_model.pth",
    repo_type="model",
)
cfg_path = hf_hub_download(
    repo_id="maelic/REACTPlusPlus_VG150",
    filename="yolo12m/config.yml",
    repo_type="model",
)

# 4. Run evaluation
import subprocess
subprocess.run([
    "python", "tools/relation_eval_hydra.py",
    "--config-path", str(cfg_path),
    "--task", "sgdet",
    "--eval-only",
    "--checkpoint", str(ckpt_path),
])

---

Citation

@article{neau2026reactpp,
  title   = {REACT++: Efficient Cross-Attention for Real-Time Scene Graph Generation
},
  author  = {Neau, Maëlic and Falomir, Zoe},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.06386},
}