CATS: Co-saliency Activated Tracklet Selection for Video Co-localization

1) Co-saliency Generation: Appropriate neighboring warped saliency maps are fused with the saliency maps of activators to generate different co-saliency maps. These maps are then similarly fused through averaging for generating eventual co-saliency object prior (O). 2) Bounding-Box Filtering: Co-saliency object prior helps in filtering out noisy bounding box proposals and keep good proposals. 3) Tracklets Generation: Tracklets are generated from good proposals up to the next activator, and are then scored using object priors at the two ends. 4) Tube Generation: Tracklets with high confidence scores forming good spatiotemporal consistency are chosen for tube generation. Goal: To localize the common object from set of similar videos, which is also known as video co-localization.