CUEMP: Correspondence Uncertainty Estimation With Motion Priors for Dense Visual Odometry

Abstract

Deep dense visual odometry has made significant advancements by leveraging dense flow fields. However, current mainstream flow-based visual odometry methods often fail to suppress the visual similarity noise in correlation volumes and rely on the inefficient four-scale pyramids for correlation sampling with non-adaptive context. These limitations make the model prone to estimating the erroneous rigid flow in ambiguous regions, which misleads the pose tracking and geometric modeling of visual odometry. We address these limitations mainly by two devised strategies. One is the learnable, motion-prior-guided correspondence uncertainty estimation with motion priors that uses predicted uncertainty to softly suppress interference within source correlation volumes. The other is the dilation-constrained deformable correlation sampling mechanism that adaptively captures robust context during correlation sampling with only two scales. Compared to existing schemes, experiments show that our visual odometry could achieve the superior performance on representative benchmarks.