Combined Holistic and Local Patches for Recovering 6d Object Pose

Abstract

We present a novel method for recovering 6D object pose in RGB-D images. By contrast with recent holistic or local patch-based method, we combine holistic patches and local patches together to fulfil this task. Our method has three stages, including holistic patch classification, local patch regression and fine 6D pose estimation. In the first stage, we apply a simple Convolutional Neural Network (CNN) to classify all the sampled holistic patches from the scene image. After that, the candidate region of target object can be segmented. In the second stage, as proposed in Doumanoglou et al. [16] and Kehl et al. [17], a Convolutional Autoencoder (CAE) is employed to extract condensed local patch feature, and coarse 6D object pose can be estimated by the regression of feature voting. Finally, we apply Particle Swarm Optimization (PSO) to refine 6D object pose. Our method is evaluated on the LINEMOD dataset [5] and the Occlusion dataset [10, 5], and compared with the state-of-the-art on the same sequences. Experimental results show that our method has high precision and good performance under foreground occlusion and background clutter conditions.