🤖 AI Summary
Existing local motion blur detection methods often rely on gradient-based shortcuts, limiting their generalization capability. To address this issue, this work introduces BOCCHI, a real-world captured benchmark in which the gradient distributions of sharp and blurred regions exhibit substantial overlap, thereby effectively mitigating shortcut learning. Furthermore, the paper proposes MSDCT-UNet, a novel architecture that integrates multi-scale Discrete Cosine Transform (DCT) priors, frequency-aware DCT attention, and FiLM modulation to achieve precise pixel-level blur localization. Trained on only 633 images, the proposed method achieves state-of-the-art performance on BOCCHI in terms of both domain-specific mIoU and boundary localization accuracy, and consistently outperforms existing approaches in cross-dataset transfer scenarios.
📝 Abstract
Local motion blur detection requires pixel-level localization of blurred regions. Existing benchmarks let models rely on gradient shortcuts that fail to transfer. We introduce BOCCHI (Blurred Objects Captured across Cameras with Human-annotated Imagery), a real-captured benchmark whose sharp regions overlap the blur gradient distribution and defeat these shortcuts, and propose MSDCT-UNet (Multi-Scale Discrete Cosine Transform UNet), a frequency-aware encoder-decoder injecting multi-scale DCT priors through DCT Attention and FiLM. MSDCT-UNet ranks first in in-domain mIoU and boundary localization on BOCCHI, and BOCCHI-trained models outperform every other training source on cross-dataset transfer with only 633 training images.