Generate

Use generate to precompute detections and embeddings that can be reused by later eval, tune, or research runs.

Examples

!!! example

=== "CLI"

    Benchmark-driven cache generation:

    ```bash
    boxmot generate --benchmark mot17 --split ablation
    ```

    Direct-source cache generation:

    ```bash
    boxmot generate \
      --source path/to/dataset \
      --detector yolov8n \
      --reid osnet_x0_25_msmt17
    ```

=== "Python"

    ```python
    from boxmot import Boxmot

    benchmark_cache = Boxmot().generate(benchmark="mot17", split="ablation")
    print(benchmark_cache.cache_dir)

    direct_cache = Boxmot(
        detector="yolov8n",
        reid="osnet_x0_25_msmt17",
    ).generate(source="path/to/dataset")
    print(direct_cache.timings["frames"])
    ```

Why generate first

Cache generation removes repeated detector and ReID work from later benchmark runs. That makes evaluation and tuning faster and more reproducible.

What gets written

generate writes cached detector outputs and ReID embeddings under the configured project/name directory so later runs can reuse them.

When to use it

• before repeated eval runs on the same benchmark
• before tune, which evaluates many tracker parameter sets
• before research, which may evaluate many candidate code variants

Public detections

Use --detection-source to cache public MOTChallenge detections instead of running a detector:

boxmot generate --benchmark mot17 --split ablation --detection-source frcnn

This downloads the public detection files from the benchmark config and generates ReID embeddings for them. Later eval and tune runs with the same --detection-source reuse this cache.

Available sources for MOT17: frcnn, sdp, dpm, or public (uses the default defined in the benchmark YAML).

See Benchmark Workflows for cache reuse, MMOT benchmark ids, and replay image-loading behavior.

CLI Arguments

::: mkdocs-click :module: boxmot.engine.cli :command: boxmot :depth: 1 :command: generate :style: table :prog_name: boxmot generate