Self-Supervised Multi-Object Tracking with Cross-Input Consistency

We propose a novel self-supervisory signal that we call cross-input consistency: we construct two distinct inputs for the same sequence of video, by hiding different information about the sequence in each input.