Geometry-aware synchronization and automatic instance segmentation for prior-free pose estimation

Develop automatic instance segmentation and geometry-aware synchronization for prior-free relative 6D pose estimation of multiple object instances, incorporating constraints such as pairwise-distance preservation to reject feature-consistent but geometrically incompatible correspondences.

Background

The PROSE method addresses prior-free relative 6D pose estimation by matching multiple co-visible instances of an unknown rigid object using fused semantic and geometric foundation features, followed by cycle-consistency refinement and RANSAC-based 3D registration. Its current formulation assumes that instance masks and a predefined anchor are already available.

The correspondence refinement enforces feature consistency and cycle consistency but does not explicitly model the rigid geometric structure of the object. Consequently, correspondences can satisfy cycle consistency while remaining geometrically incorrect, particularly when descriptors are ambiguous or repetitive. The unresolved research direction is therefore to remove the reliance on predefined segmentation and to design synchronization methods that use geometric constraints, such as preservation of pairwise distances, to reject mutually consistent but physically incompatible matches.