Related Experiment Video
Updated: Jun 14, 2025
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
2.6K
Category-Level Object Pose Estimation with Statistic Attention
Changhong Jiang1, Xiaoqiao Mu2, Bingbing Zhang3
1School of Electrical and Electronic Engineering, Changchun University of Technology, Changchun 130012, China.
Sensors (Basel, Switzerland)
|August 29, 2024
Summary
SAPENet improves six-dimensional object pose estimation by incorporating statistical attention for higher-order feature analysis. This method enhances the understanding of complex shapes and object differences, achieving state-of-the-art results.
Area of Science:
- Computer Vision
- 3D Object Pose Estimation
Background:
- Category-level object pose estimation using 3D-GC shows promise but struggles with long-range dependencies and detailed object differences.
- Existing methods using self-attention or Transformers focus on first-order features, neglecting complex information and fine-grained distinctions.
Purpose of the Study:
- To propose SAPENet, an enhanced 3D-GC architecture for improved six-dimensional object pose estimation.
- To address limitations in capturing long-range dependencies and discerning subtle object variations.
Main Methods:
- Replaced 3D-GC in the encoder with an HS-layer for feature extraction.
- Incorporated statistical attention to compute higher-order statistical information.
- Developed specialized sub-modules for pose regression, point cloud reconstruction, and bounding box voting, with statistical attention in pose regression.
Main Results:
- Achieved a mean Average Precision (mAP) of 49.5 on the 5°2 cm metric, a 3.4-point improvement over the baseline.
- Demonstrated state-of-the-art (SOTA) performance on the REAL275 dataset.
Conclusions:
- SAPENet effectively captures higher-order statistical information and geometric relationships, outperforming existing methods.
- The proposed approach significantly advances category-level object pose estimation, particularly for complex and occluded objects.
