Vide-omics: A genomics-inspired paradigm for video analysis

Profile Hidden Markov Models for Foreground Object Modelling

“…The proposed method was evaluated on the Berkeley Motion Segmentation Dataset (BMS-26) [7], which is a widely used benchmark for motion segmentation. Similarly to [14], twelve videos with moving cameras were selected: people2, cars1-10 (PTZ motion) and marple10 (freely moving camera). The foreground outputs produced by the vide-omics inspired algorithm for foreground extraction [14] were used as reference, since that approach has a low false positive rate.…”

“…Similarly to [14], twelve videos with moving cameras were selected: people2, cars1-10 (PTZ motion) and marple10 (freely moving camera). The foreground outputs produced by the vide-omics inspired algorithm for foreground extraction [14] were used as reference, since that approach has a low false positive rate. The performance of the proposed foreground enhancement method was compared using the F1 score against the GrowCut algorithm (GC) [22], a variational method (Ochs) [16] and a CRF based method for image segmentation (CRF) [15].…”

“…The performance of the proposed foreground enhancement method was compared using the F1 score against the GrowCut algorithm (GC) [22], a variational method (Ochs) [16] and a CRF based method for image segmentation (CRF) [15]. In order to create the P-HMMs, the UPGMA algorithm was selected to produce the phylogenetic trees [20], where the inter-segment distances were calculated using pair-wise sequence alignment [14] and then normalised [12]. Note that to ensure foreground segments are discriminative enough, very short ones, i.e.…”

“…Eventually, foreground pixels identified by P-HMM search were added to the initial foreground mask. Finally, as in [14], some post-processing was applied: small regions (area lower than the square of 1% of the frame's width) were removed.…”

“…In order to address some of the challenges encountered by visual surveillance systems, including camera motion, a novel video analysis paradigm, 'vide-omics', has recently been proposed [14]. Inspired by the principles of genomics, this paradigm interprets videos as sets of temporal measurements of a scene in constant evolution without setting any constraint in terms of camera motion, object behaviour or scene structure.…”