Recognizing Objects And Reasoning About Their Interactions

Abstract

The task of scene understanding involves recognizing the different objects present in the scene, segmenting the scene into meaningful regions, as well as obtaining a holistic understanding of the activities taking place in the scene. Each of these problems has received considerable interest within the computer vision community. We present contributions to two aspects of visual scene understanding.

First we explore multiple methods of feature selection for the problem of object detection. We demonstrate the use of Principal Component Analysis to detect avifauna in field observation videos. We improve on existing approaches by making robust decisions based on regional features and by a feature selection strategy that chooses different features in different parts of the image. We then demonstrate the use of Partial Least Squares to detect vehicles in aerial and satellite imagery. We propose two new feature sets; Color Probability Maps are used to capture the color statistics of vehicles and their surroundings, and Pairs of Pixels are used to capture captures the structural characteristics of objects. A powerful feature selection analysis based on Partial Least Squares is employed to deal with the resulting high dimensional feature space (almost 70,000 dimensions). We also propose an Incremental Multiple Kernel Learning (IMKL) scheme to detect vehicles in a traffic surveillance scenario. Obtaining task and scene specific datasets of visual categories is far more tedious than obtaining a generic dataset of the same classes. Our IMKL approach initializes on a generic training database and then tunes itself to the classification task at hand.

Second, we develop a video understanding system for scene elements, such as bus stops, crosswalks, and intersections, that are characterized more by qualitative activities and geometry than by intrinsic appearance. The domain models for scene elements are not learned from a corpus of video, but instead, naturally elicited by humans, and represented as probabilistic logic rules within a Markov Logic Network framework. Human elicited models, however, represent object interactions as they occur in the 3D world rather than describing their appearance projection in some specific 2D image plane. We bridge this gap by recovering qualitative scene geometry to analyze object interactions in the 3D world and then reasoning about scene geometry, occlusions and common sense domain knowledge using a set of meta-rules.