Supplementary material

Approach Overview

Input Image		Extract superpixels

Obtain several spatiotemporal closure hypotheses using parametric maxflow		Build a graph with spatial and temporal affinities

Superpixel Extraction

No temporal information	With temporal information

The figures show a single superpixel tracked by following the temporal connections in the superpixel graph.
Using temporal information results in a more temporally stable superpixel segmentation (right) then if we don't use it (left).

Results

We test our algorithm on several videos. For each video, we show the original video and several solutions of interest.
For the car video we also show that shots can be detected by detecting minimas in global frame affinity.

Video	Solutions