MSSPA-GC: Multi-Scale Shape Prior Adaptation with 3D Graph Convolutions for Category-Level Object Pose Estimation

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Category-level object pose estimation predicts 6D pose and size for objects within known categories.
• Large intra-class shape variations pose a significant challenge.
• Shape prior adaptation in normalized canonical coordinates (NOCS) helps mitigate variation, but existing methods struggle with complex geometries.

Purpose of the Study:

• To propose a novel shape prior adaptation method, MSSPA-GC, for improved category-level object pose estimation.
• To address limitations of existing methods in handling complex object structures and intra-class shape variations.
• To achieve state-of-the-art performance with improved efficiency.

Main Methods:

• Inputting observed instance point clouds and prior shape point clouds into a novel network.
• Utilizing a 3D graph convolution network for pose-aware features and a PointNet-like MLP for shape-aware features.
• Aggregating features via multi-scale propagation for comprehensive 3D object descriptors.

Main Results:

• Achieved state-of-the-art performance on REAL275 and CAMERA25 datasets.
• Demonstrated superior performance on objects with complex geometric structures.
• Required only 25% of the parameters compared to existing shape prior adaptation models.
• Showcased good generalization ability on the REDWOOD75 dataset.

Conclusions:

• The proposed MSSPA-GC method effectively enhances category-level object pose estimation.
• The novel feature aggregation strategy leads to robust and accurate pose prediction.
• The method offers a more parameter-efficient and generalizable solution for 6D object pose estimation.