Oriented Box Tracking

Use OBB tracking when object angle matters, such as aerial imagery or rotated documents.

boxmot track \
  --detector yolo11n-obb.pt \
  --geometry obb \
  --tracker ocsort \
  --source video.mp4 \
  --save

Canonical contract

import torch

from boxmot.structures import Detections, OrientedBoxes

detections = Detections(
    geometry=OrientedBoxes(cxcywha_float32_cpu.contiguous()),
    scores=scores_float32_cpu.contiguous(),
    class_ids=class_ids_int64_cpu.contiguous(),
    sample_id="sequence-1:42",
)
tracks = tracker.update(detections)

OrientedBoxes contains float32[N,5] (cx, cy, w, h, angle) values. Width and height are positive, angles are radians, and finite angles remain unwrapped for temporal continuity. Tracker instances have a fixed AABB or OBB mode.

Before updates, OBB trackers resolve equivalent rectangle representations relative to the track state and retain damped angular motion. Association uses oriented geometry in OBB mode.

For a standalone box-only call, tracker.update(rows) also accepts an exact NumPy N x 7 matrix in (cx, cy, w, h, angle, confidence, class_id) order. The return value is a C-contiguous float64 M x 9 matrix in (cx, cy, w, h, angle, track_id, confidence, class_id, detection_index) order. Use Detections for Tracks, enrichments, sample metadata, and pipelines.

Review association configuration and native support before comparing implementations.