Video segmentation by detection for the 2019 Unsupervised DAVIS Challenge

In this work, we present a new framework, video segmentation by detection (VSD), for tackling the problem of unsupervised video multi-object segmentation. Our model employs an object detector for automatic target discovery and a set of single-object trackers for the simultaneous tracking of all targets. While addressing the object re-identification problem, we observe that many of the objects of interest in the dataset are humans or human centric such as bicycles. As such, following a design philosophy that special purpose algorithms will always be better than general purpose ones, we explore whether we can leverage the rich existing research efforts on re-identifying humans to improve the results or exploit the spatial relations of human-centric objects to humans. The proposed method achieves the highest J -Mean of 0.535 and an overall second place in the unsupervised track of the 2019 DAVIS Challenge