Deep Learning for Canonical Reconstruction of Deformable Objects from Depth Images in Robotic Vision and SLAM-Aware Perception

Abstract

Reconstructing the canonical pose of non-rigid objects from arbitrary depth observations is an important problem in robotic vision, particularly for systems that must perceive, track, and interact with deformable objects in dynamic environments. In robotics and SLAM-related perception, depth cameras are widely used to support object recognition, spatial understanding, scene mapping, and motion analysis. However, non-rigid deformation remains challenging because the same object may appear in significantly different poses, making reliable object-level representation and tracking difficult. In this study, we present a deep learning approach for reconstructing the default canonical pose of non-rigid objects from single depth images. The proposed model combines short-range and long-range feature extraction with the original depth input to capture both local geometric details and global structural information. By transforming arbitrary posed observations into a consistent canonical representation, the method supports more stable shape understanding, pose normalization, and object-level perception for robotic systems operating in real-world environments. This is particularly relevant to robotic vision tasks involving human motion analysis, deformable-object tracking, manipulation, and semantic mapping. The model is trained on synthetic human datasets and evaluated on synthetic human, real human, and animal datasets. Experimental results demonstrate improved retrieval accuracy compared with existing methods, showing that the proposed approach can generalize across different non-rigid categories and sensing conditions. These findings highlight the potential of canonical pose reconstruction as a useful component for intelligent robotic perception, depth-based scene interpretation, and SLAM-aware systems that require robust understanding of deformable objects.