Rapid Forensic Search & Retrieval in Video Archives
Re-id is a challenging problem for several reasons. Camera views are non-overlapping so conventional tracking methods are not applicable. View angles, illumination and calibration for the two cameras are generally arbitrary, leading to significant variation in appearance, to the point that features seen in one camera are often missing in the other. Consequently, face recognition methods or those based on matching visual features are often unreliable.
While re-id has received significant interest, much of this effort can be viewed from the perspective of multi-class classification, namely, methods that seek to classify each probe image into one of gallery images. Within this framework, re-id literature can be categorized into two themes with one theme focusing on cleverly designing local features and the other focusing on metric learning.
In contrast, we developed a new conceptual framework that is fundamentally different from the existing literature. We view re-id as an instance of bipartite graph matching. We simultaneously match all or a sub-collection of probes to the gallery images. This is natural for many surveillance contexts, such as in airports, where multiple entities are viewed in a camera at any time. Figure 1 illustrates re-id with two camera views as a weighted bipartite matching problem, where images from the two views are taken as nodes in the graph, edges link nodes from the two views, and weights are associated with edges.
Students Currently Involved in Project
- Yuting Chen
- Gregory Castanon