MSDCT-UNet — Pretrained Weights

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

File	Train source	In-domain mIoU	Cross-domain mIoU
bocchi.pth	BOCCHI	0.850	0.403
reloblur.pth	ReLoBlur	0.930	0.388
blurdataset.pth	BlurDataset	0.640	0.276

All three are the same architecture (MultiTransformUNet, alias DCTEverywhereUNet, paper name MSDCT-UNet) trained on different sources. Each is 113 MB (FP32, model_state only — optimizer state stripped).

Companion artifacts

• Pre-built PKLs (HiFST DCT features + img + mask): https://huggingface.co/datasets/aianonymous12/BOCCHI
• Code release (training, inference, eval scripts): see the anonymous repo URL in the paper's supplementary.

Quick download

pip install huggingface_hub
huggingface-cli download aianonymous12/msdct-unet \
    bocchi.pth reloblur.pth blurdataset.pth \
    --local-dir checkpoints

Verify (after cloning the code release)

python eval.py --config configs/bocchi.yaml --ckpt checkpoints/bocchi.pth --split val
# Expected: mean_iou ≈ 0.850 (within ±0.005)

License

CC BY-NC 4.0 — non-commercial research use with attribution.

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.