Omni6DPose: A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking

摘要

"6D object pose estimation is crucial in the field of computer vision. However, it suffers from a significant lack of large-scale and diverse datasets, impeding comprehensive model evaluation and curtailing downstream applications. To address these issues, this paper introduces , a substantial benchmark featured by its diversity in object categories, large scale, and variety in object materials. is divided into three main components: (Real 6D Object Pose Estimation Dataset), which includes images annotated with over annotations across instances in categories; (Simulated 6D Object Pose Estimation Dataset), a simulated training set created by mixed reality and physics-based depth simulation; and (Pose Aligned 3D Models), the manually aligned real scanned objects used in and . is inherently challenging due to the substantial variations and ambiguities. To address this issue, we introduce , an enhanced version of the SOTA category-level 6D object pose estimation framework, incorporating two pivotal improvements: Semantic-aware feature extraction and Clustering-based aggregation. Moreover, we provide a comprehensive benchmarking analysis to evaluate the performance of previous methods on this new large-scale dataset in the realms of 6D object pose estimation and pose tracking."