Browse Teach Curate

Hadamard-Attention Linear Transformer for Stereo Matching

An efficient linear-time stereo-matching transformer (HART) that replaces quadratic self-attention with a Hadamard-product attention, adds a Dense Attention Kernel that removes the upper bound on attention weights to escape the low-rank bottleneck, and a Multi-Kernel and Order Interaction module unifying semantic and spatial features, improving disparity estimation in reflective and weakly textured regions.

Data Text

Highlights Hadamard Attention Recurrent Transformer: A Strong Baseline for Stereo Matching

Introduces HART (Hadamard Attention Recurrent Stereo Transformer), an efficient linear-time transformer for stereo matching - computing per-pixel correspondences (disparity, hence depth) between two …