WAVE-Stereo: Warp-Aligned Volume Encoding for Stereo Matching

WAVE-Stereo addresses the long-standing dichotomy in iterative stereo matching: methods that rely on explicit correlation volumes (which preserve matching ambiguity but are computationally heavy) versus those that use feature warping (which are efficient for local refinement but discard global matching context). The authors propose a unified architecture that integrates these two paradigms to achieve a superior balance of accuracy and speed.

The core of the framework is the GeoWarp Correspondence Encoder (GWCE). Instead of choosing between correlation search or warping, GWCE processes three parallel inputs: correlation features from a geometry encoding volume, the current disparity estimate, and warped right-image features. By fusing these, the model simultaneously leverages explicit matching candidate information and local cross-view residuals. To overcome the limited receptive field of the iterative ConvGRU updates, the authors introduce Periodic Global Context Propagation (PGCP). This mechanism periodically injects global spatial information into the iterative process, which is particularly effective for resolving ambiguities in textureless or repetitive regions.

By avoiding the heavy 3D convolutional cost aggregation and the high latency of external foundation models, WAVE-Stereo achieves real-time inference at 66ms. It demonstrates competitive zero-shot generalization across five major benchmarks, including KITTI 2015 and Middlebury, proving that a well-designed architecture with unified correspondence representation can outperform more complex, resource-intensive models.