Point matching as a classification problem for fast and robust object pose estimation

Abstract

We propose a novel approach to point matching under large viewpoint and illumination changes that is suitable for accurate object pose estimation at a much lower computational cost than state-of-the-art methods. Most of these methods rely either on using ad hoc local descriptors or on estimating local affine deformations. By contrast, we treat wide baseline matching of keypoints as a classification problem, in which each class corresponds to the set of all possible views of such a point. Given one or more images of a target object, we train the system by synthesizing a large number of views of individual keypoints and by using statistical classification tools to produce a compact description of this view set. At run-time, we rely on this description to decide to which class, if any, an observed feature belongs. This formulation allows us to use powerful and fast classification methods to reduce matching error rates. In the context of pose estimation, we present experimental results for both planar and non-planar objects in the presence of occlusions, illumination changes, and cluttered backgrounds. We will show that our method is both reliable and suitable for initializing real-time applications.

Training Binary Descriptors for Improved Robustness and Efficiency in Real-Time Matching

“…Lepetit et al [11] revolutionized the keypoint matching problem by casting it as a classification problem. Given an image of a target object, a feature setF is generated by combining the statistics of the set of warped patches, referred to as a viewset, around each keypoint.…”

“…Lepetit and Fua [12] further extended their initial approach in [11] by replacing the NN classifier with a random forest. Image patches were recognized on the basis of very simple, randomly chosen binary tests which were grouped into decision trees which then recursively partitioned the space of all possible viewsets.…”

“…To avoid the use of a priori information of camera intrinsics, most approaches [11][12][13][14][15] apply the affine transformation to synthesize these images. We step away from this convention and apply a perspective transformation instead.…”

“…Invariance to various deformations can be achieved by synthesizing a sufficient amount of viewsets. PCA was applied in [11] to extract a compact description for all the viewsets associated with each keypoint. At runtime, a nearest-neighbour (NN) classifier was then employed to identify keypoints which are inF .…”

“…We present a model generation framework which can be applied to any descriptor-based approach, allowing us to achieve realtime matching. We strictly focus on objects which are planar, thus allowing us to synthesize images of the object in order to capture the appearance of keypoint patches under several perspectives [11][12][13][14].…”