Signal, Image and Video Processing Обработка сигналов

SyDE-Net: synergistic disparity estimation network for stereo matching

Yan, RongHuan, Tian, Weixin, Matthew, Kaggwa
Springer Nature · 2026-11-17 · DOI 10.1007/s11760-026-05873-7

Аннотация

Stereo matching continues to face challenges in accuracy and robustness, particularly in ill-posed regions such as texture-less areas, repetitive textures, and occlusions. To address these issues, we propose the Synergistic Disparity Estimation Network (SyDE-Net). The network is a hybrid monocular-stereo framework that integrates global structural priors from monocular depth estimation with geometric constraints from binocular stereo matching. Within this framework, we design a Semantic-Geometric Decoupled Attention (SGDA) module. This module uses Fourier-Gated Residual Channel Attention to recalibrate monocular feature channels and self-attention to aggregate spatial context in the stereo branch. To strengthen collaboration between the two branches during refinement, we introduce an asymmetric residual guidance strategy that combines cross-branch residual-difference encoding with branch-dependent input-group ordering. Guided by this strategy, multi-scale Convolutional Gated Recurrent Units (ConvGRUs) iteratively refine the disparity maps. Extensive experiments demonstrate that SyDE-Net achieves strong performance on Scene Flow and Stereo Correspondence and Reconstruction of Endoscopic Data (SCARED), validating its effectiveness. On Scene Flow, our model achieves an end-point error (EPE) of 0.370 pixels, with 4.06% of pixels having absolute disparity errors exceeding one pixel (Bad-1), yielding the lowest errors among the compared methods. Regional evaluation further shows improved accuracy across all nine low-texture mask configurations. After fine-tuning on SCARED, it achieves a mean absolute depth error (MAE) of $$3.582 \pm 0.007$$ 3.582 ± 0.007 mm across four runs with different random seeds (mean ± sample standard deviation). The source code for SyDE-Net is publicly available at https://github.com/lingshijiang/SyDE-Net .

Читать у издателя →