In this paper, we describe an efficient method for creating mosaics from collections of videos. Our method is based on a utility maximization formulation which we optimize greedily. We employ a function for quickly estimating mosaics without computing image features that allows us to efficiently take greedy steps while still achieving user definable goals for mosaic quality. Indeed, we demonstrate using a number of single- and multi-video experiments that our approach can construct high-quality mosaics in only a fraction of the time required to perform the operations undertaken by existing video mosaicing algorithms. While we focus in this work on the application of panorama construction, our method has a wide range of applications, such as super-resolution, summary, and indexing.

This work presents a discriminative training method for particle filters in the context of multi-object tracking. We are motivated by the difficulty of hand-tuning the many model parameters for such applications and also by results in many application domains indicating that discriminative training is often superior to generative training methods. Our learning approach is tightly integrated into the actual inference process of the filter and attempts to directly optimize the filter parameters in response to observed errors. We present experimental results in the challenging domain of American football where our filter is trained to track all 22 players throughout football plays. The training method is shown to significantly improve performance of the tracker and to significantly outperform two recent particle-based multi-object tracking methods.

The task of registering video frames with a static model is a common problem in many computer vision domains. The standard approach to registration involves finding point correspondences between the video and the model and using those correspondences to numerically determine registration transforms. Current methods locate video-to-model point correspondences by assembling a set of reference images to represent the model and then detecting and matching invariant local image features between the video frames and the set of reference images. These methods work well when all video frames can be guaranteed to contain a sufficient number of distinctive visual features. However, as we demonstrate, these methods are prone to severe misregistration errors in domains where many video frames lack distinctive image features. To overcome these errors, we introduce a concept of local distinctiveness which allows us to find model matches for nearly all video features, regardless of their distinctiveness on a global scale. We present results from the American football domain-where many video frames lack distinctive image features-which show a drastic improvement in registration accuracy over current methods. In addition, we introduce a simple, empirical stability test that allows our method to be fully automated. Finally, we present a registration dataset from the American football domain we hope can be used as a benchmarking tool for registration methods.

For many multi-part object classes, the set of parts can vary not only in location but also in type. For example, player formations in American football involve various subsets of player types, and the spatial constraints among players depend largely upon which subset of player types constitutes the formation. In this work, we study the problem of localizing and classifying the parts of such objects. Pictorial structures provide an efficient and robust mechanism for localizing object parts. Unfortunately, these models assume that each object instance involves the same set of parts, making it difficult to apply them directly in our setting. With this motivation, we introduce the mixture-of-parts pictorial structure (MoPPS) model, which is characterized by three components: a set of available parts, a set of constraints that specify legal part subsets, and a function that returns a pictorial structure for any legal part subset. MoPPS inference corresponds to jointly computing the most likely subset of parts and their positions. We propose a restricted, but useful, representation for MoPPS models that facilitates inference via branch-and-bound optimization, which we show is efficient in practice. Experiments in the challenging domain of American football show the effectiveness of the model and inference procedure.