Disney Research


Re-identification refers to the task of finding the same subject across a network of surveillance cameras. This task must deal with appearance changes caused by variations in illumination, a person’s pose, camera viewing angle and background clutter. State-of-the-art approaches usually focus either on feature modeling – designing image descriptors that are robust to changes in imaging conditions, or dissimilarity functions – learning effective metrics to compare images from different cameras. Typically, with novel deep architectures both approaches can be merged into a single end-to-end training, but to become effective, this requires annotating thousands of subjects in each camera pair. Unlike standard CNN-based approaches, we introduce a spatial pyramid-like structure to the image and learn CNNs for image sub-regions at different scales. When training a CNN using only image sub-regions, we force the model to recognize not only the person’s identity but also the spatial location of the sub-region. This results in highly effective feature representations, which when combined with Mahalanobis-like metric learning significantly outperform state-of-the-art approaches.

Copyright Notice

The documents contained in these directories are included by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a non-commercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author’s copyright. These works may not be reposted without the explicit permission of the copyright holder.