No abstract

It is too often that tracking algorithms lose track of interest points in image sequences. This persistent problem is difficult because the pixels around an interest point change in appearance or move in unpredictable ways. In this paper we explore how classifying videos into categories of camera motion improves the tracking of interest points, by selecting the right specialist motion model for each video. As a proof of concept, we enumerate a small set of simple categories of camera motion and implement their corresponding specialized motion models. We evaluate the strategy of predicting the most appropriate motion model for each test sequence. Within the framework of a standard Bayesian tracking formulation, we compare this strategy to two standard motion models. Our tests on challenging real-world sequences show a significant improvement in tracking robustness, achieved with different kinds of supervision at training time.frame 0 frame 10 frame 20 Figure 1: Example sequence with overlaid box showing the output of our specialized "forward" motion model, where the velocity and scale of objects approaching the camera tend to increase. Neither Brownian nor constant-velocity motion models are as successful at tracking interest points here.

No abstract

Training object detectors aims at choosing specific visual attributes which are efficient and optimal for each learned object. This paper presents a new process which achieves this goal by putting all families of descriptors one wants to consider in a pool of descriptors and by letting the algorithm build a cascade with the most efficient descriptors by introducing management of very large features pools. On the one hand, the selection of a specific family of descriptors for a given application implies a deep experience of the operator on the algorithm behaviour. On the other hand, physical constraints such as computer time and memory requirements prevent us from using all available descriptors. We present here a solution which considers several families of descriptors as a pool of descriptors and builds a cascade with the most efficient descriptors. The idea developed here consists in beginning to build a cascade with one type of descriptors and then introducing new kinds of descriptors when the current descriptor family does not bring enough differentiating information anymore. In this scope, four families of descriptors are studied here: Histogram Distance on Haar Region (HDHR), Edge Orientation Histograms (EOH), Histogram Orientation Gradient (HOG) and Gabor filters. Evaluation on public data sets shows the importance of complementary features, since performances of state of the art methods are improved.

Multiclass classification is the core issue of many pattern recognition tasks. In some applications, not only the predicted class is important but also the confidence associated to the decision. This paper presents a complete framework for multiclass classification that recovers probability estimates for each class. It focuses on the automatic configuration of the system so that no user-provided tuning is needed. No assumption about the nature of data or the number of classes is done either, resulting in a generic system. A suitable decomposition of the original multiclass problem into several biclass problems is automatically learnt from data. State-of-the-art biclass classifiers are optimized and their reliabilities are assessed and considered in the combination of the biclass predictions. Quantitative evaluations on different datasets show that the automatic decomposition and the reliability assessment of our system improve the classification rate compared to other schemes, as well as it provides probability estimates of each class. Besides, it simplifies considerably the user effort to use the framework in a specific problem, since it adapts automatically.