Re-Identification, or the task of finding the same person across a network of surveillance cameras, requires annotating thousands of subjects in each camera pair, explain the writers of a paper written by two Disney researchers, Slawomir Bak and Peter Carr for the IEEE International Conference on Advanced Video and Signal-Based Surveillance (AVSS) 2017. 

The researchers propose introducing what they call a “spatial pyramid-like structure” to the image, forcing the model to recognize the person’s identity and the spatial location of the sub-regions at different scales. The results, they say, are highly effective feature representations and outperform state-of-the-art approaches. 

To see the paper, visit www.disneyresearch.com/publication/deep-spatial-pyramid/.