Point cloud registration is crucial in computer vision and robotics, yet it struggles with ambiguous structures and low-overlap regions. This paper proposes GenReg-Amb, a generative registration framework using flow-based models to address these challenges. It models registration as a conditional generation task, learning a velocity field for point transformation, and introduces an attention mechanism for handling structural ambiguity and an overlap-aware sampling strategy for consistent region generation.
Key findings
GenReg-Amb improves registration recall by up to 18% on low-overlap scenes.
The framework enables uncertainty quantification and multiple hypothesis generation.
It achieves state-of-the-art performance on benchmarks like 3DMatch and 3DLoMatch.
Limitations & open questions
The computational efficiency of the model in real-time applications is a potential area for improvement.