Vehicle Disambiguation from Multiple Observations

摘要

We introduce disambiguation as a generalization of re-identification. The field of re-identification has been used in identifying objects like people and vehicles passing by surveillance cameras. However, all of these problems have one thing in common: they all begin with a query image and seek to find matching images in galleries from other cameras. In our problem, a database is provided of objects captured from multiple cameras, but a query image is not provided. The challenge then is to disambiguate the suspect by searching through the detections across cameras and identifying the common object. In this paper we introduce the problem of disambiguation and we show our multi-stage approach using deep convolutional neural networks and a pair of reasoning engines which identify a common target even without a known query image to reference. For the purposes of this paper, we select vehicles as the target dataset, but we believe that our approach translates directly to other object classes (e.g. people).