BOCCHI Motion-Blur Detection

Companion dataset for the ACCV 2026 submission "MSDCT-UNet: Multi-Scale DCT U-Net for Local Motion Blur Detection".

This repo ships both raw image+mask pairs and pre-computed PKL feature caches, so you can either train from scratch or skip preprocessing entirely.

Code release: https://huggingface.co/aianonymous12/msdct-unet (pretrained weights for the same architecture).

---

Repository layout

aianonymous12/BOCCHI/
├── README.md
├── BOCCHI_dataset/                              ← OUR dataset (BOCCHI, CC BY-NC 4.0)
│   ├── img/   0001.jpg ... 0633.jpg          # 633 RGB images
│   ├── mask/  0001.png ... 0633.png          # 633 binary masks (0 = sharp, 255 = blur)
│   └── all_data.pkl                          # pre-built feature cache (8.4 GB)
├── Inference_dataset/merged/                  ← Cross-eval set (mixed sources)
│   ├── img_all/    422 jpg                   # 164 BOCCHI-protocol + 162 ReLoBlur test
│   ├── mask_all/   422 png                   #   + 96 BlurDataset held-out
│   └── all_data.pkl                          # pre-built (5.6 GB)
├── ReLoBlur_dataset/train/
│   └── all_data.pkl                          # pre-built only (16 GB) — third-party
└── BlurDataset/
    └── all_data.pkl                          # pre-built only (2.7 GB) — third-party

Subset	Samples	Raw size	PKL size
BOCCHI (BOCCHI_dataset)	633	109 MB	8.4 GB
Inference (merged)	422	97 MB	5.6 GB
ReLoBlur (train)	1200	—	16 GB
BlurDataset	200	—	2.7 GB

Raw img+mask for BOCCHI and Inference are bundled (CC BY-NC 4.0). ReLoBlur and BlurDataset raw data are not included (third-party licenses); only the derived PKL features are redistributed.

---

Quick start

pip install huggingface_hub

# Minimum: BOCCHI + Inference PKLs (14 GB) — reproduces Table 2 BOCCHI + Table 3
huggingface-cli download aianonymous12/BOCCHI --repo-type dataset \
    BOCCHI_dataset/all_data.pkl \
    Inference_dataset/merged/all_data.pkl \
    --local-dir data

# Raw img+mask only (205 MB, no PKLs) — for building your own features
huggingface-cli download aianonymous12/BOCCHI --repo-type dataset \
    --include "BOCCHI_dataset/img/*" "BOCCHI_dataset/mask/*" \
              "Inference_dataset/merged/img_all/*" "Inference_dataset/merged/mask_all/*" \
    --local-dir data

# Everything (~33 GB)
huggingface-cli download aianonymous12/BOCCHI --repo-type dataset --local-dir data

Then follow §4 of the code release README to reproduce paper Tables 2 / 3.

---

PKL internal schema

{
    "<stem>": {
        "img":      np.ndarray (H, W, 3)      uint8,          # RGB
        "msk":      np.ndarray (H, W)         uint8 {0,255},  # 0 = sharp, 255 = blur
        "DCT_coef": np.ndarray (Wgrid, Hgrid, 57)  float32,   # HiFST DCT features
    },
    "settings": {"mode": "rotate"/"pad", "size": (720, 1080),
                 "num_scales": 4, "scale_start": 2, ...}
}

The 57-channel layout is the legacy (Wgrid, Hgrid, 57) format; the loader in the code release (utils/dataset.py) auto-detects this and permutes to (57, Hgrid, Wgrid) PyTorch convention at __getitem__ time.

---

License

• BOCCHI (BOCCHI_dataset/img/, BOCCHI_dataset/mask/, BOCCHI_dataset/all_data.pkl) and the BOCCHI-protocol portion of the Inference set: CC BY-NC 4.0 (our own data, free for non-commercial research use with attribution).
• ReLoBlur PKL: derivative of Li et al., AAAI 2023. Redistributed here under the original authors' terms for review purposes. Please cite the original ReLoBlur paper if you use it.
• BlurDataset PKL: derivative of the CUHK blur-detection benchmark. Same conditions as above.

The reviewer / reproducer must comply with each source dataset's license when using the corresponding subset.

---

Citation

@inproceedings{anon2026msdctunet,
  title  = {MSDCT-UNet: Multi-Scale DCT U-Net for Local Motion Blur Detection},
  author = {Anonymous},
  booktitle = {ACCV},
  year   = {2026}
}

Note

Hosted on an anonymous reviewer account for the ACCV 2026 double-blind review. Author identities will be revealed at camera-ready.