A Topic-Motion Model for Unsupervised Video Object Discovery.pdf

A Topic-Motion Model for Unsupervised Video Object Discovery David Liu and Tsuhan Chen Department of Electrical and Computer Engineering Carnegie Mellon University /projects/DISCOV Abstract The bag-of-words representation has attracted a lot of attention recently in the field of object recognition. Based on the bag-of-words representation, topic models such as Probabilistic Latent Semantic Analysis (PLSA) have been applied to unsupervised object discovery in still images. In this paper, we extend topic models from still images to motion videos with the integration of a temporal model. We propose a novel spatial-temporal framework that uses topic models for appearance modeling, and the Probabilis- tic Data Association (PDA) filter for motion modeling. The spatial and temporal models are tightly integrated so that motion ambiguities can be resolved by appearance, and ap- pearance ambiguities can be resolved by motion. We show promising results that cannot be achieved by appearance or motion modeling alone. 1. Introduction Discovering objects in video is a challenging task. By discovering, we mean that the object can be a person, a car, or a building. Without having any prior knowledge about the object type or its position, we would like to identify an object from a video that occurs over a period of time. This is particularly challenging when the image sequence has low resolution and consists of highly cluttered background. This is not easily achieved by directly applying motion-based or unsupervised appearance-based methods in literature; see Figure 1. Some methods observe the same scene over a long time and build a color distribution model for each pixel [25] [11] [19]. Unusual objects can then be identified if some pix- els observe substantial deviation from their long-term color distribution models. These kind of background modeling approaches are suitable for video surveillance with a static camera, but if an image sequence is obtained from a mov- ing camera, the