Transformer and Trainable Bilateral Filter for Unsupervised Stereo Matching

Recent advances have demonstrated that depth estimation from stereo image pairs can be formulated as an unsupervised learning task solved using convolutional neural networks (CNNs). However, existing architectures exhibit two critical limitations: insufficient capacity for global feature extraction, and suboptimal cost-aggregation performance under noisy conditions. To address these challenges, we propose TBSMnet (Transform-Bilateral Stereo Matching Network), a novel architecture comprising two key components: a hybrid ViT-CNN encoder that leverages the complementary strengths of vision transformers and CNNs to capture long-range dependencies and a trainable bilateral filtering module that suppresses noise in homogeneous regions yet preserves geometric boundaries during aggregation, significantly improving precision. Comprehensive evaluations on benchmark datasets demonstrate the effectiveness of our approach.