Joint Depth Map Super-Resolution Method via Deep Hybrid-Cross Guidance Filter
Ke Wang,Lijun Zhao,Jinjing Zhang,Jialong Zhang,Anhong Wang,Huihui Bai
DOI: https://doi.org/10.1016/j.patcog.2022.109260
IF: 8
2022-12-17
Pattern Recognition
Abstract:Nowadays color-guided Depth map Super-Resolution (DSR) methods mainly have three thorny problems: (1) joint DSR methods have serious detail and structure loss at very high sampling rate; (2) existing DSR networks have high computational complexity; (3) color-depth inconsistency makes it hard to fuse dual-modality features. To resolve these problems, we propose a joint hybrid-cross guidance filter method to progressively recover the quality of degraded Low-Resolution (LR) depth maps by exploiting color-depth consistency from multiple perspectives. Specifically, the proposed method leverages pyramid structure to extract multi-scale features from High-Resolution (HR) color image. At each scale, hybrid side window filter block is proposed to achieve high-efficiency color feature extraction after each down-sampling for HR color image. This block is also used to extract depth features from the LR depth map. Meanwhile, we propose a multi-perspective cross-guided fusion filter block to progressively fuse high-quality multi-scale structure information of color image with corresponding enhanced depth features. In this filter block, two kinds of space-aware group-compensation modules are introduced to capture various spatial features from different perspectives. Meanwhile, color-depth cross-attention module is proposed to extract color-depth consistency features for impactful boundary preservation. Comprehensively qualitative and quantitative experimental results have demonstrated that our method can achieve superior performances against a lot of state-of-the-art depth SR approaches in terms of mean absolute deviation and root mean square error on Middlebury, NYU-v2 and RGB-D-D datasets.
computer science, artificial intelligence,engineering, electrical & electronic