2018/12/24 by Kyaw Zaw Lin, Lin, Kyaw Zaw, Weipeng Xu +7
Computer Science · Engineering · #3D Shape Modeling and Analysis #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Human Pose and Action Recognition #Image Processing and 3D Reconstruction
paper · pdf · doi:10.48550/arxiv.1812.09899
openalex publication_date 2018/12/24 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
We propose a novel approach to jointly perform 3D shape retrieval and pose\nestimation from monocular images.In order to make the method robust to\nreal-world image variations, e.g. complex textures and backgrounds, we learn an\nembedding space from 3D data that only includes the relevant information,\nnamely the shape and pose. Our approach explicitly disentangles a shape vector\nand a pose vector, which alleviates both pose bias for 3D shape retrieval and\ncategorical bias for pose estimation. We then train a CNN to map the images to\nthis embedding space, and then retrieve the closest 3D shape from the database\nand estimate the 6D pose of the object. Our method achieves 10.3 median error\nfor pose estimation and 0.592 top-1-accuracy for category agnostic 3D object\nretrieval on the Pascal3D+ dataset, outperforming the previous state-of-the-art\nmethods on both tasks.\n