Skip to content

Compare BoxMOT Components

Choose components against the constraints of your footage and deployment, then compare them on the same cached detections. Tracker and ReID rankings are not portable across detectors, datasets, hardware, or association settings.

Choose a tracker

Requirement Good starting point Why
Small, fast motion-only baseline ByteTrack or SFSORT No appearance model is required.
Abrupt or non-linear motion OcSort Observation-centric updates reduce motion-model drift.
Moving camera BotSort or DeepOcSort Camera-motion compensation is available.
Crowds and longer occlusions StrongSort, DeepOcSort, HybridSort, BoostTrack, or OccluBoost Appearance features can reconnect identities.
Instance masks and current images are available MafHda Combines AABB motion with masked correlation-filter appearance; requires nonempty full-frame masks.
2D and 3D detections with camera calibration EagerMOT Tracks in 3D, uses 2D observations during 3D detection gaps, and supports optional ego poses.
In-process native C++ BotSort, ByteTrack, OcSort, OccluBoost, or SFSORT These trackers have registered native live backends; cached workflows stream keyed rows through that same API.

This is a starting-point guide, not a universal ranking. The tracker overview contains the support matrix, and each tracker page explains its own requirements and tradeoffs.

Compare ReID models

Appearance-assisted trackers trade more compute for another identity cue. Use the ReID model table to compare parameters, compute, embedding size, and available retrieval results. Compare runtime formats on your target device as well: PyTorch, ONNX, TensorRT, OpenVINO, Core ML, and TFLite can have different latency and deployment constraints.

Run a fair tracker comparison

For image trackers, materialize detections and embeddings once, then reuse the same immutable build for each tracker:

boxmot materialize --experiment mot17/ablation-yolox-lmbn.yaml

boxmot eval --experiment mot17/ablation-yolox-lmbn.yaml --build BUILD_ID --tracker bytetrack
boxmot eval --experiment mot17/ablation-yolox-lmbn.yaml --build BUILD_ID --tracker botsort
boxmot eval --experiment mot17/ablation-yolox-lmbn.yaml --build BUILD_ID --tracker boosttrack

Keep the experiment, detector profile, split, postprocessing, and metric configuration fixed. Record the tracker YAML and runtime overrides with every result.

For EagerMOT, use eval --dataset with the same multimodal dataset. Keep 2D/3D predictions, calibration, ego poses, and scoring protocol fixed across runs.

Read the metrics together

  • HOTA balances detection, association, and localization quality.
  • IDF1 emphasizes identity consistency.
  • MOTA combines misses, false positives, and identity switches.
  • FPS and latency must be measured on the deployment hardware and include any ReID or camera-motion stages you enable.

See Evaluation and Postprocessing for metric details and Experiment Workflows for cache identity and reproducibility.