Research On Multi-view Object Class Detection

Recently, multi-view object class detection has received more and more attention in the fields such as computer vision and multimedia. And it has extended to much more general object class rather than face and pedestrian. The goal of multi-view object detection is to detect any instance of this object class at arbitrary pose or viewpoint. Instances of the same object class can vary greatly in color, shape and texture, and even more greatly across viewpoints of the same object class. Multi-view object class detection has to deal with both inter-class and across viewpoint variations. As far, there are some efficient algorithms for object class detection. But models dealing with viewpoint variations have not been studied as much, mainly due to the fact that it is difficult to find accurate correspondences across viewpoints of an object class. Therefore, building models considering both inter-class and across viewpoints variations is in need but more challenging.This paper focuses on models for multi-view object class detection andproposed two methods for building correspondences between multiple viewpoints of the object class. For the first method, we train a separate classifier for each viewpoint for detecting instances of this specific viewpoint and then arrange the classifier on a view sphere reflecting the underlying intrinsic relationships between different viewpoints. Then we triangulate the sphere into a number of uniformly distributed meshes to represents the explicit correspondences across view detectors. As a result, multi-view objects from untrained viewpoints can be detected by combining the outputs of the adjacent view detectors on the sphere. For the second method, we build a visual codebook which contains patches describing local shape or contour structure for each viewpoint. Then we build correspondences across viewpoints by link the visual codebooks from different viewpoints together. The links are obtained by matching similar patches across instances of the object class but from different viewpoints. Finally we build a multi-view codebook for object detecting from arbitrary viewpoints. We evaluate our methods on several public datasets and show the effectiveness compared to other state-of-the-art multi-view methods.The main contributions of this paper are as follows:1) We propose a novel triangular combining scheme based on the view sphere. Instances from arbitrary viewpoint are detected by combining the outputs of the view detectors using the triangular scheme.2) The single view visual codebook of the object class is organized into a decision-tree structure. The construction of the tree corresponds to the generating of the visual codebook. This strcture can efficiently reduce the matching computation while testing. The matched codebook entries cast hough votes and the local maximas of the hough images are regarded as object hypothese.3) We build across viewpoint correspondences by low-level similar patch matching, integrating useful information from adjacent viewpoints at an earlier phase, forming a multi-view codebook for detection of multi-view object instances.