Skip to content

Mode Defaults

Shared tracking-workflow defaults live in boxmot/configs/runtime.yaml.

What it controls

runtime.yaml centralizes defaults for:

  • shared detector and ReID selections
  • runtime options such as imgsz, device, batch_size, and sequence_workers
  • command-specific defaults for track, materialize, eval, tune, and research

Current shape

shared:
  detector: yolov8n
  reid: osnet_x0_25_msmt17

runtime:
  tracker: bytetrack
  tracker_backend: python
  sequence_workers: auto
  save: false

research:
  proposal_model: openai/gpt-5.4
  max_metric_calls: 24
  eval_timeout: 900.0

The engine CLI resolves tracking defaults through this file. Reusable Python components instead receive explicit immutable specs. ReID training defaults remain in boxmot/reid/training/configs/defaults.yaml, while export defaults remain in boxmot/reid/exporters/defaults.yaml.

sequence_workers: auto selects automatic replay parallelism for evaluation and tuning: min(selected_sequences, max(1, logical_cpus - 2)) workers. A positive integer supplies an explicit cap, bounded by the selected sequence count. In tuning, this count applies separately to each trial. See sequence parallelism for visualization constraints.