DeepSORT
DeepSORT
DeepSORT (Wojke et al., 2017) was the breakthrough that added a deep appearance descriptor to the original SORT tracker, dramatically reducing identity switches under occlusion. A small CNN trained on Market-1501 produces 128-D embeddings; cosine distance plus Mahalanobis distance from the Kalman filter drive matching. While newer trackers (ByteTrack, BoT-SORT,…
Definition
DeepSORT (Wojke et al., 2017) was the breakthrough that added a deep appearance descriptor to the original SORT tracker, dramatically reducing identity switches under occlusion. A small CNN trained on Market-1501 produces 128-D embeddings; cosine distance plus Mahalanobis distance from the Kalman filter drive matching. While newer trackers (ByteTrack, BoT-SORT, OC-SORT) outperform DeepSORT on benchmarks, it remains widely deployed because of its simplicity and the wealth of open-source integrations. FI Tech still uses DeepSORT in legacy edge boxes where Hailo-8 cannot easily run the larger BoT-SORT re-ID head.
