3D object detection: integrating sparse encoders with transformer-based decoders

Abstract

Accurate 3D object detection is essential for precise robot manipulation and as sembly tasks in industrial environments, where reliable object classification and localization significantly influence the efficiency and robustness of automated systems. Although existing point cloud encoders such as SECOND, PointPillars, and PV-RCNN have demonstrated strong feature extraction capabilities for sparse and dense 3D point clouds, they generally lack effective mechanisms to capture comprehensive global spatial relationships between objects. To address this limitation, this dissertation proposes the integration of specialized transformer based decoders with encoders, combining their strengths in local geometric en coding and global feature context aggregation.

Experiments conducted on three widely used 3D open source datasets, KITTI and nuScenes, showcasing the effectiveness of our Voxel-DETR3D architecture. The results show promising results in evaluation metrics in all datasets when using task-specific decoders, achieving notable performance gains.

These findings demonstrate that the proposed integration of encoder and decoder substantially enhances object detection accuracy, facilitating improvements in the reliability and effectiveness of robotic assembly tasks.

Keywords: Transformer Decoder, 3D Object Detection, Point Cloud, Voxeliza tion, Industrial Assembly.