Unusual event detection algorithm

Unusual events are the video segments that have correspondences to distinctive important features, and vice versa. Using the co-embedding coordinate $x$ computed in (7), we define inferred similarity among the video segments $v_i,v_j$ as: $S_{x}(v_i,v_j) = e^{- \vert\vert x(v_i) - x(v_j) \vert\vert _2}$, as seen in figure 5(f). In the co-embedding space, detecting unusual video segments is done by finding spatially isolated clusters. We choose the following simple solution for this:

• We apply K-means algorithm on the embedding coordinates $x$ to cluster video segments into disjoint sets $\{VC_1, \ldots, VC_c\}$,
• and we compute inter-cluster similarity on $S_{VC}(VC_i,VC_j) = \sum_{v_i \in VC_i, v_j \in VC_j}S_{x}(v_i,v_j)$.
• The clusters $VC_k$, with small total inter-cluster similarity value, $\sum_{i}S_{VC}(VC_i, VC_k) < \theta$, are considered as unusual events.