Detectors

Detector configs live under boxmot/configs/detectors and define reusable models independently of datasets and experiments.

id: yolox-x-mot17

box_type: aabb

classes:
  0: person

inference:
  image_size: [800, 1440]
  confidence_threshold: 0.01

checkpoints:
  ablation:
    path: models/yolox_x_MOT17_ablation.pt
    uri: https://...
  test:
    path: models/yolox_x_MOT17_test.pt
    uri: https://...

Experiments select checkpoints explicitly. A dataset split never implicitly chooses a detector checkpoint.

The resolver requires exactly two positive image_size values in height-width order, a confidence threshold in [0, 1], and matching detector/dataset box types.

Experiments resolve their detector profile before computing a build ID. Materialization accepts only an experiment and has no detector or geometry override; create another experiment to select a different detector checkpoint. Python callers construct a DetectorSpec with the resolved artifact path, SHA-256, preprocessing, precision, geometry, and normalized backend options.