Skip to content
← Models
Open weightsByteDance / HUST/10 trustVerified

ByteTrack

Multi-object tracking by associating every detection box, including the low-confidence ones everyone else throws away.

Visit official page →

About

ByteTrack's insight is simple: most trackers discard low-confidence detections, but those boxes often contain real occluded objects. BYTE associates nearly every detection box, matching high-score boxes to tracklets first, then using low-score boxes to recover true objects and filter background. Paired with a plain YOLOX detector and a Kalman motion model, with no ReID network and no attention, it set a new state of the art.

The authors report 80.3 MOTA, 77.3 IDF1, and 63.1 HOTA on MOT17 at 30 FPS on a single V100, plus SOTA on MOT20, HiEve, and BDD100K. MIT-licensed and dependency-light, ByteTrack became the default tracking baseline in production systems and the association method ported into countless trackers. If your tracker keeps losing people behind each other, this is usually the first fix to try.

Specs

Context window

n/a

Pricing

Free (open weights, MIT)

Free quota

Free — self-hosted; official code and pretrained models on GitHub

API access

No official API; GitHub code and checkpoints; Roboflow hosted inference

Weights

Open weights

Released

Not disclosed

Benchmarks

Author-reported: 80.3 MOTA / 77.3 IDF1 / 63.1 HOTA on MOT17 test at 30 FPS on V100; SOTA on MOT20, HiEve, BDD100K

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

multi-object-trackingvideo-analyticssurveillancesports-analytics

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: ECCV 2022 paper (arXiv:2110.06864) with MOT17/MOT20 numbers confirmed via the official repo README; MIT license confirmed via multiple third-party attributions and the Roboflow model page. Strengths: deceptively simple idea with huge practical impact, permissive license, widely deployed. Docked for: (1) tracking quality is bounded by the underlying detector; (2) no appearance model, so long occlusions still break identity. License: MIT.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
9.0/10
Maintenance · recency, releases, issue load
6.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · ECCV 2022 (arXiv 2110.06864, Oct 2021)
  • · 80.3 MOTA / 77.3 IDF1 / 63.1 HOTA on MOT17 at 30 FPS on V100, author-reported
  • · BYTE association ported into many production trackers

Sources