Disney’s High-Efficiency AI Motion-Tracking Technology

This CVPR 2026 paper presents a memory- and compute-efficient sampler for all-pairs correlation volumes in optical-flow estimation. It preserves RAFT’s mathematical operator while targeting faster, hi

Optical flow, which estimates pixel movement between two video frames, serves as the basis for various video-processing tasks, including frame interpolation, video compression, and inpainting. This paper focuses on sampling all-pairs correlation volumes used by the RAFT family. Existing approaches faced a tradeoff: precomputing values for every pixel combination caused memory requirements to grow rapidly, while calculating the required values on demand slowed execution.

Karlis Martins Briedis, Markus Gross, and Christopher Schroers analyzed the fact that the actual sampling process references only small, regular portions of the volume rather than the entire volume. Based on this observation, they proposed a partial-block sparse correlation volume that computes only the necessary blocks, along with a patch-centered memory layout. The paper explains that this method preserves the exact mathematical operations defined by RAFT.

In evaluations, the proposed method achieved up to 92% higher performance than on-demand sampling at the same low memory usage, while reducing memory usage by up to 99% compared with the baseline implementation. Total inference time for high-resolution inputs could be reduced by up to 63%. The researchers examined high-resolution accuracy and execution-time performance using an 8K dataset based on the BLENDER film CHARGE and an extension of SEA-RAFT’s inference stage.

Maru