Nowadays, object recognition in 3D scenes has become an emerging challenge with various applications. However, an object can’t be represented well by artificial features derived only from 2D images or depth images separately, and supervised learning method usually requires lots of manually labeled data. To address those limitations, we propose a cross-modality deep learning framework based on contractive autoencoders for 3D scenes object recognition. In particular, we use contractive autoencoding to learn feature representations from 2D and depth images at the same time in an unsupervised way, it is possible to capture their joint information to reinforce detector training. Experiments on 3D image dataset demonstrate the effectiveness of the proposed method for 3D scene object recognition.

Date Published: 29 August 2016
PDF: 7 pages
Proc. SPIE 10033, Eighth International Conference on Digital Image Processing (ICDIP 2016), 100330S (29 August 2016); doi: 10.1117/12.2243988

Published in SPIE Proceedings Vol. 10033:
Eighth International Conference on Digital Image Processing (ICDIP 2016)
Charles M. Falco; Xudong Jiang, Editor(s)