RL-ScanIQA

RL-ScanIQA, published in the CVPR 2026 proceedings, formulates blind 360° image quality assessment as active perception. A PPO-trained policy learns where to look, while an attention-based quality assessor learns how those selected viewports explain perceptual quality. Both components are optimized jointly from IQA supervision, without requiring human scanpath annotations.

• Task-driven scanpaths — viewport selection is optimized for quality assessment instead of imitating gaze trajectories.
• Multi-level rewards — step-wise exploration, scanpath-set diversity, and task-aligned perceptual feedback stabilize policy learning.
• Cross-domain robustness — distortion-space augmentation and rank-consistent losses improve transfer across 360° IQA datasets.
• Research-ready release — reusable PyTorch modules, a runnable reference trainer, paper settings, tests, and curated qualitative results.

flowchart LR
    A["360° ERP image"] --> B["32 candidate viewports<br/>8 yaw × 4 pitch"]
    A --> C["Frozen DINOv2<br/>global feature"]
    B --> D["DINOv2<br/>viewport features"]
    C --> E["GRU policy + value head"]
    D --> E
    E --> F["K scanpaths × T steps"]
    F --> G["Attention-based<br/>quality assessor"]
    G --> H["Predicted MOS"]
    G -. "quality-driven reward" .-> E

The paper setting uses 32 candidates, a 90° × 90° field of view, 224 × 224 viewports, and averages K=15 scanpaths of length T=7 at inference.

These values are transcribed from the paper's in-dataset table. The release does not bundle trained checkpoints or raw experiment logs, so they are paper-reported results rather than claims reproduced by this checkout. Cross-dataset results and ablations are collected in docs/RESULTS.md.

git clone https://github.com/wangyuji1/RLScanIQA.git
cd RLScanIQA

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .

For development and tests:

python -m pip install -e ".[dev]"
pytest -q

The reference trainer can generate a tiny synthetic pair dataset when --pairs_csv is empty. The following command validates the end-to-end code path on CPU; it is not a paper-scale experiment:

python scripts/train_reference.py \
  --pairs_csv '' \
  --epochs 1 \
  --batch_size 8 \
  --device cpu \
  --n_yaw 2 \
  --n_pitch 1 \
  --K 1 \
  --T 1 \
  --viewport_hw 32 \
  --d_f 32 \
  --d_h 32

For the paper-scale geometry and inference settings, start from:

python scripts/train_reference.py \
  --pairs_csv /path/to/pairs.csv \
  --img_root /path/to/erp_images \
  --epochs 300 \
  --batch_size 4 \
  --device cuda \
  --backbone dino \
  --K 15 \
  --T 7

The DINOv2 option downloads model code and weights through torch.hub on first use. See docs/REPRODUCIBILITY.md before running full experiments.

Raw CVIQD, OIQA, and JUFE data are intentionally not redistributed. The reference trainer expects pair supervision:

img1,img2,Q1,Q2
images/example_a.jpg,images/example_b.jpg,64.67,16.15

Paths are resolved relative to --img_root. A malformed CSV or a missing image raises an error instead of silently substituting synthetic data. See examples/pairs.example.csv and docs/DATASETS.md.

RLScanIQA/
├── rl_scaniqa/              # Reusable policy, PPO, rewards, losses, and geometry
├── scripts/
│   └── train_reference.py   # Runnable reference training path
├── tests/                   # CPU unit and smoke tests
├── assets/                  # Curated paper/supplementary visualizations
├── docs/                    # Results, datasets, and reproducibility notes
├── examples/                # Input-format examples
├── CITATION.cff
└── pyproject.toml

If this work helps your research, please cite:

@InProceedings{Wang_2026_CVPR,
  author    = {Wang, Yujia and Li, Yuyan and Liu, Jiuming and Zhang, Fang-Lue and Zheng, Xinhu and Dodgson, Neil A.},
  title     = {RL-ScanIQA: Reinforcement-Learned Scanpaths for Blind 360deg Image Quality Assessment},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2026},
  pages     = {37401--37412}
}

Machine-readable metadata is available in CITATION.cff.

• This repository does not include the benchmark datasets, private review material, pretrained checkpoints, or personal documents.
• The standalone trainer is a readable reference implementation. Exact paper reproduction additionally depends on the authors' dataset preprocessing, split files, DINOv2 setup, and training infrastructure.
• No open-source license has been selected yet. Until the authors choose one, standard copyright applies; please contact the authors before reuse or redistribution.

This work was supported by the Marsden Fund Council, managed by the Royal Society of New Zealand, under Grant MFP-20-VUW-180.