WHTMix: Efficient Stereo Depth Estimation via Walsh-Hadamard Token Mixing

Name: WHTMix: Efficient Stereo Depth Estimation via Walsh-Hadamard Token Mixing
Uploaded: 2026-10-17T04:58:00.213221+00:00
Duration: 712 s
Description: The authors replaced slow global self-attention in stereo transformers with a fast Walsh-Hadamard token mixer. This data-independent mixer mixes tokens globally in the transform domain at log-linear cost, while keeping data-dependent cross-attention for correspondence. On synthetic driving data, this change reduced model compute by 2.46x and inference latency by 2.65x while maintaining accuracy.

The authors replaced slow global self-attention in stereo transformers with a fast Walsh-Hadamard token mixer. This data-independent mixer mixes tokens globally in the transform domain at log-linear cost, while keeping data-dependent cross-attention for correspondence. On synthetic driving data, this change reduced model compute by 2.46x and inference latency by 2.65x while maintaining accuracy.

Open the full episode