README.md

---
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: 10.52
            name: mR@20
          - type: R@20
            value: 18.32
            name: R@20
          - type: F1@20
            value: 13.36
            name: F1@20
          - type: mR@50
            value: 13.22
            name: mR@50
          - type: R@50
            value: 22.54
            name: R@50
          - type: F1@50
            value: 16.67
            name: F1@50
          - type: mR@100
            value: 13.96
            name: mR@100
          - type: R@100
            value: 23.77
            name: R@100
          - type: F1@100
            value: 17.59
            name: F1@100
          - type: e2e_latency_ms
            value: 19.4
            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: 10.32
            name: mR@20
          - type: R@20
            value: 20.0
            name: R@20
          - type: mR@50
            value: 13.94
            name: mR@50
          - type: R@50
            value: 26.9
            name: R@50
          - type: mR@100
            value: 16.48
            name: mR@100
          - type: R@100
            value: 32.08
            name: R@100
          - type: mean_recall
            value: 21.87
            name: mean_recall
  - name: REACT++ yolov8m
    results:
      - task:
          type: object-detection
          name: Scene Graph Detection
        dataset:
          name: VG150
          type: vg150
        metrics:
          - type: mR@20
            value: 12.05
            name: mR@20
          - type: R@20
            value: 22.78
            name: R@20
          - type: F1@20
            value: 15.76
            name: F1@20
          - type: mR@50
            value: 15.42
            name: mR@50
          - type: R@50
            value: 28.73
            name: R@50
          - type: F1@50
            value: 20.07
            name: F1@50
          - type: mR@100
            value: 16.51
            name: mR@100
          - type: R@100
            value: 30.84
            name: R@100
          - type: F1@100
            value: 21.51
            name: F1@100
          - type: e2e_latency_ms
            value: 17.8
            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](https://github.com/Maelic/SGG-Benchmark) framework and described in the
[REACT++ paper (Neau et al., 2026)](https://arxiv.org/abs/2603.06386).

---

## Results — SGDet on VG150 test split (ONNX, CUDA)

> Metrics from end-to-end ONNX evaluation (`tools/eval_onnx_psg.py`). E2E Latency = image load + pre-process + ONNX forward.

| Backbone | Params | R@20 | R@50 | R@100 | mR@20 | mR@50 | mR@100 | F1@20 | F1@50 | F1@100 | E2E Lat. (ms) |
|----------|:------:|-----:|-----:|------:|------:|------:|-------:|------:|------:|-------:|--------------:|
| yolo12m | ~20.2M | 18.32 | 22.54 | 23.77 | 10.52 | 13.22 | 13.96 | 13.36 | 16.67 | 17.59 | 19.4 |
| yolo26m | ~20.2M | 20.0 | 26.9 | 32.08 | 10.32 | 13.94 | 16.48 | - | - | - | - |
| yolov8m | ~25.9M | 22.78 | 28.73 | 30.84 | 12.05 | 15.42 | 16.51 | 15.76 | 20.07 | 21.51 | 17.8 |

---

## Checkpoints

| Variant | Sub-folder | Checkpoint files |
|---------|------------|-----------------|
| yolo12m | `yolo12m/` | `yolo12m/model.onnx` (ONNX) · `yolo12m/best_model_epoch_19.pth` (PyTorch) |
| yolo26m | `yolo26m/` | `yolo26m/react_pp_yolo26m.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)

```python
from huggingface_hub import hf_hub_download

onnx_path = hf_hub_download(
    repo_id="maelic/REACT-pp-VG150",
    filename="yolo12m/react_pp_yolo12m.onnx",
    repo_type="model",
)
# Run with tools/eval_onnx_psg.py or load directly via onnxruntime
```

### PyTorch

```python
# 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/REACT-pp-VG150",
    filename="yolo12m/best_model.pth",
    repo_type="model",
)
cfg_path = hf_hub_download(
    repo_id="maelic/REACT-pp-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

```bibtex
@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},
}
```