Track, Check, Repeat: An EM Approach to Unsupervised Tracking
This is the official code release for our CVPR21 paper on unsupervised detection and tracking. It produces results slightly better than reported in the paper.
[Paper] [Project Page]
We use ensemble agreement between 2d and 3d models, as well as motion cues, to unsupervisedly learn object detectors from scratch. Top: 3d detections. Middle: 2d segmentation. Bottom-left: unprojected 2d segmentation, in a bird’s eye view. Bottom-right: 3d detections, in a bird’s eye view.
Overview
An EM approach to unsupervised tracking. We present an expectation-maximization (EM) method, which takes RGBD
videos as input, and produces object detectors and trackers as output. (a) We begin with a handcrafted E step, which uses optical