Bingham procrustean alignment for object detection in clutter

Abstract

A new system for object detection in cluttered RGB-D images is presented. Our main contribution is a new method called Bingham Procrustean Alignment (BPA) to align models with the scene. BPA uses point correspondences between oriented features to derive a probability distribution over possible model poses. The orientation component of this distribution, conditioned on the position, is shown to be a Bingham distribution. This result also applies to the classic problem of least-squares alignment of point sets, when point features are orientation-less, and gives a principled, probabilistic way to measure pose uncertainty in the rigid alignment problem. Our detection system leverages BPA to achieve more reliable object detections in clutter.

Sliding Shapes for 3D Object Detection in Depth Images

“…Keypoint matching: Just as SIFT keypoints for image matching [50], a popular type of algorithms [51][52][53][54][55][56][57][58][59][60] is to detect keypoints on a 3D point cloud or a mesh, generate descriptors for the keypoints (e.g. spin image and 3D shape context), and use the matching to align with models in the training data.…”

Data association for semantic world modeling from partial views

“…The fitness function is computed as a weighted linear combination of least-squares range-image errors and point-cloud nearest-neighbor distances (approximated with a distance transform of the model point cloud). The detection pipeline is related to the system described by Glover and Popovic (2013).…”

Unsupervised object individuation from RGB-D image sequences

“…In dynamic situations, the estimation of O k i can be predicted by using filtering-based tracking or robust 3D registration methods such as [1], [20], [11], [15], [3]. In this research, GMM-based robust 3-d registration with Gaussian Sum Filtering (GSF) method is used, which is proposed in the authors' previous work in [5], with considering many outliers which are points belonging to another object as in the contact case of Fig.…”

“…In order to implement the process from visual sensor data, previous research broadly falls into two categories: individuationby-location and individuation-by-feature. Individuation-bylocation identifies each object by referring to their locations in one image [16], [10] or in sequential images [1], [20], while individuation-by-feature utilizes general features such as edge or color differences to characterize each object [3], [15] or distinguish each by using specific feature information defined in the object database [8], [2].…”

“…First, the unstructured condition prohibits supervised methods in that a robot cannot obtain ground-truth information of unknown target objects in advance. For example, the approach needs prior knowledge of target objects as specified features [3], pre-defined models [1], [20], [15] and off-line learning [2]. On the other hand, unsupervised methods suffer from the dynamic condition in robustness against dynamic interactions of multiple objects.…”