Multi-Scale Feature Extraction Network with Half-Wavelet Pyramid Attention Mechanism for Robust 6D Pose Estimation

D object pose estimation is very important in robotics and augmented reality domains. However, existing RGBbased methods face significant challenges in terms of blocking robustness, diverse feature acquisition, and multi-scale feature extraction. Traditional attention mechanisms fail to fully exploit frequency domain features. Existing multi-scale extraction modules have limited performance. To address these problems, we propose a new end-to-end framework that integrates HalfWavelet Pyramid Attention module(HWPA) and Multi-Scale Feature Extraction module(MSFE), which are embedded in ResNet34 as the backbone network. HWPA collects and fuses normal domain and wavelet domain features simultaneously through a hierarchical pyramid structure. MSFE adopts differentiated receptive field design to effectively extract multi-scale features of objects. Experiments on LM and LM-O show that our method achieves 94.46 % and 67.03 % accuracy under ADD(-S) metric, respectively. Our study provides a solution for occlusion robustness and establishes a new normal form of normal-frequency domain feature acquisition, which has important application value for industrial automation and AR navigation systems.