Search papers, labs, and topics across Lattice.
The paper introduces Freq-DP Net, a novel dual-branch neural network for removing fence occlusions from single images captured using dual-pixel (DP) sensors. It leverages defocus disparity, modeled with a cost volume, and the fence's global structural pattern, learned via Fast Fourier Convolution (FFC), as complementary geometric and structural priors. Experiments on a newly created benchmark dataset demonstrate that Freq-DP Net achieves state-of-the-art performance compared to existing general-purpose methods for single-image fence removal.
Dual-pixel sensors unlock surprisingly effective single-image fence removal, outperforming existing methods by fusing geometric and frequency-based priors.
Removing fence occlusions from single images is a challenging task that degrades visual quality and limits downstream computer vision applications. Existing methods often fail on static scenes or require motion cues from multiple frames. To overcome these limitations, we introduce the first framework to leverage dual-pixel (DP) sensors for this problem. We propose Freq-DP Net, a novel dual-branch network that fuses two complementary priors: a geometric prior from defocus disparity, modeled using an explicit cost volume, and a structural prior of the fence's global pattern, learned via Fast Fourier Convolution (FFC). An attention mechanism intelligently merges these cues for highly accurate fence segmentation. To validate our approach, we build and release a diverse benchmark with different fence varieties. Experiments demonstrate that our method significantly outperforms strong general-purpose baselines, establishing a new state-of-the-art for single-image, DP-based fence removal.
