Answers are grounded in this paper and cite their sources.
Deep learning methods have achieved remarkable progress in fringe projection depth estimation. However, existing parallel CNN-Transformer architectures still suffer from high computational complexity, as well as over-smoothing of high-frequency fringe details and distortion of low-frequency global structures caused by feature inconsistencies during the interaction between the two branches. To address these limitations, we propose a frequency-decoupled parallel CNN-Mamba network, PCMNet, in which the CNN branch primarily extracts local high-frequency details, while the Mamba branch efficiently models long-range spatial dependencies with linear complexity via 2D selective scan. A frequency-decoupled bidirectional feature interaction (FDBFI) module is designed to decouple the dual-branch features through the discrete wavelet transform and enhance high-frequency details while avoiding distortion of global structures. In addition, a wavelet and depthwise convolution based cross-feature multi-scale fusion (WDCFMS) module is designed to aggregate multi-scale features, achieve complementary fusion of the dual-branch features, and suppress reconstruction artifacts. Experiments on two fringe projection datasets demonstrate that PCMNet with only 17.94 M parameters achieves optimal results on most metrics while also delivering the best visual results.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: