Reproduce “Rethinking RAFT for Efficient Optical Flow” on OpenResearch

About this paper

Navid Eslami, Farnoosh Arefi, Amir M. Mansourian, Shohreh Kasaei

Researchers at Sharif University of Technology developed Ef-RAFT, an enhancement to the RAFT optical flow framework, to improve estimation in challenging scenarios like large displacements and repetitive patterns. The method achieved over a 10% reduction in End-Point Error (EPE) on the Sintel dataset and nearly a 6% improvement in the F1-All metric on the KITTI-15 dataset, while maintaining a balanced trade-off in computational efficiency.

Key insights

• Dynamically adjusting the query distribution in correlation volume lookups can effectively improve the network's ability to handle large displacements and reduce data redundancy.
• Augmenting local features with global, attention-derived positional context helps resolve ambiguities in repetitive or untextured regions, leading to more robust optical flow estimation.
• Significant accuracy gains in optical flow can be achieved by targeted enhancements to core architectural components without incurring disproportionately high computational overhead, enabling practical application on standard hardware.

Summary by alphaXiv.