Zhang Jiyao ; Wu Mingdong ; Dong Hao

Abstract

Object pose estimation plays a vital role in embodied AI and computer vision,enabling intelligent agents to comprehend and interact with their surroundings.Despite the practicality of category-level pose estimation, current approachesencounter challenges with partially observed point clouds, known as themultihypothesis issue. In this study, we propose a novel solution by reframingcategorylevel object pose estimation as conditional generative modeling,departing from traditional point-to-point regression. Leveraging score-baseddiffusion models, we estimate object poses by sampling candidates from thediffusion model and aggregating them through a two-step process: filtering outoutliers via likelihood estimation and subsequently mean-pooling the remainingcandidates. To avoid the costly integration process when estimating thelikelihood, we introduce an alternative method that trains an energy-basedmodel from the original score-based model, enabling end-to-end likelihoodestimation. Our approach achieves state-of-the-art performance on the REAL275dataset, surpassing 50% and 60% on strict 5d2cm and 5d5cm metrics,respectively. Furthermore, our method demonstrates strong generalizability tonovel categories sharing similar symmetric properties without fine-tuning andcan readily adapt to object pose tracking tasks, yielding comparable results tothe current state-of-the-art baselines.