Boosting accuracy for efficient 2D stereo matching via attention-guided fusion and spatially adaptive modulation

Description

Zhong Wu, Lili He, Juan Li, and Wenhuan Wu present a 2026 vision network for depth from paired cameras. Matching pixels between left and right views gives depth. The most exact methods lean on heavy math. Small devices cannot afford that cost, so light models often lose accuracy.

Their design swaps the heavy step for cheaper layered volumes. Flat filters process those volumes instead of bulky ones. A learned map then reweights the depth guesses by place in the scene. A fusion step borrows context from low detail views. That step makes up for the thin stack of layers. The authors report depth scores and speed that close much of the gap. Gains like these suit robots, cars, and medical scans.

Catalog record

Source: https://europepmc.org/article/MED/PMC13578262