deep-sort: Simple Online and Realtime Tracking with a Deep Association Metric

September 2019

tl;dr: use a combination of appearance metric and bbox for tracking.

Overall impression

The problem with sort is the frequent ID switches as sort uses a simple motion model and does not handle occluded tracks well.

Deep sort uses the appearance features to track objects through longer periods of occlusion. In reality the cost only consists of appearance metrics, although bbox distance is used as a gating process.

deepSORT extracts features from image patches, while PointTrack extracts features from 2D point cloud.

Key ideas

Technical details


Code available at this repo.