Towards Space-Time Semantics in Two Frames

We present a novel, low-level scheme to analyze spatial and temporal change within a local support region. Assuming available region correspondences between two adjacent frames, we divide each region into a regular grid of patches. Depending on the change of an image function inside the patch over time, each patch is assigned weights for the following four labels: “C” for a constant patch, “O” when new information originates from outside the support region, “I” for “inner” changes, and “N” for information from neighboring patches. Our method goes beyond optical flow, as it provides an additional semantic level of understanding the changes in space-time. We demonstrate how our novel “COIN” scheme can be used to categorize local space-time events in image pairs, including locally planar support regions, 3D discontinuities, and virtual vs. real crossings of 3D structures.