Integration of Probabilistic Pose Estimates from Multiple Views
We propose an approach to multi-view object detection and pose estimation that considers combinations of single-view estimates. It can be used with most existing single-view pose estimation systems, and can produce improved results even if the individual
- PDF / 3,477,885 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 51 Downloads / 201 Views
Abstract. We propose an approach to multi-view object detection and pose estimation that considers combinations of single-view estimates. It can be used with most existing single-view pose estimation systems, and can produce improved results even if the individual pose estimates are incoherent. The method is introduced in the context of an existing, probabilistic, view-based detection and pose estimation method (PAPE), which we here extend to incorporate diverse attributes of the scene. We tested the multiview approach with RGB-D cameras in different environments containing several cluttered test scenes and various textured and textureless objects. The results show that the accuracies of object detection and pose estimation increase significantly over single-view PAPE and over other multiple-view integration methods. Keywords: Pose estimation
1
· Object recognition · Multiple cameras
Introduction
Detection and pose estimation of textureless objects are well-studied challenges in robot vision. However, there are still problems that need to be solved. One of the problems is that the estimated pose can be ambiguous due to the ambiguity in the detected shape of the object [22] as shown in Fig. 1a. When a probabilistic, appearance-based pose-estimation method is used, it can be difficult to determine the viewing angle of the object due to similar appearances from the observed views. Another problem is due to the presence of outliers [9] (Fig. 1b). One of the solutions to overcome these difficulties is to observe the scene with multiple cameras. To use multiple attributes of the scene would also improve the pose estimation performance. In this paper, we introduce an approach that uses RGBD images from different viewpoints to overcome these difficulties. Multi-view integration can face difficult problems when the objects are occluded or totally unseen in one of the views as shown in Fig. 1c. Another difficulty can arise when the sensor information is incomplete or noisy. Noise or incompleteness may even result from interference between multiple RGB-D cameras as shown in Fig. 2. Therefore, we consider the integration of information from multiple RGB-D cameras and pose estimation in the presence of noisy or incomplete data as a coupled problem. c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part VII, LNCS 9911, pp. 154–170, 2016. DOI: 10.1007/978-3-319-46478-7 10
Integration of Probabilistic Pose Estimates from Multiple Views
(a)
155
(b)
(c)
Fig. 1. Some of the problems that can be solved with integration of multi-view pose estimations. (a) Ambiguities in the pose of an object; (b) Correct pose estimates are shown with green bounding boxes in two views. Outliers, which are shown with red, are eliminated after integration; (c) The cup is not visible in the right view. The integration method is capable of finding the object even if it is not visible in all of the views. The images are taken from the MPII Multi-Kinect Dataset [20]. (Color figure online)
For each view, possible 6DoF (3DoF in transla
Data Loading...
