Format `seconds` as `H:MM:SS` or `M:SS`.

Associate detections to tracks based on IOU. Args: iou_matrix: IOU cost matrix. direction_consistency_matrix

Associate detections to tracks based on Similarity (IoU) using the Jonker-Volgenant algorithm approach with no initialization instead

Wrap a scalar value in a list, pass lists through, and return None as-is. Enum members are converted to their ``.value`` so callers always re

Extract parameter-to-description mapping from Google-style Args section. Supports multiple formats including `param: desc`, `param (type): desc`,

Return validated asset entries for a single split.

Validate and convert a dataset identifier to a ``Dataset`` enum.

Return validated split names, defaulting to all available splits.

Create a test frame with deterministic pixel values for verification. Each frame has all pixels set to the same value derived from the index.

Return the expected pixel value for a frame created with create_frame(index).

Get observation from delta_t steps ago. Looks back up to delta_t timesteps in the observation history. Falls back to the most recent

Load from a JSON file. Args: path: Source file path. Returns: `BenchmarkResult` instance. Raises:

* Strips Python REPL prompts (>>> and ...) from code text. * Also removes output lines (lines that don't start with >>> or ...). * * NOTE: This is

Format as a table string. Args: columns: Metric columns to include. If `None`, includes all available metrics.

Process new detections and assign track IDs. Matches incoming detections to existing tracks, creates new tracks for unmatched detecti

Calculate IoU or IoA between two sets of boxes. Args: boxes1: First set of boxes with shape `(N, 4)`. boxes2: Second set of boxes

Clamp geometric components based on active state representation.

Compute MD5 hex digest of a file.

Detect good features to track in the grayscale image. Args: grayscale: Grayscale image. Returns: Array of fe

Generate label strings for each detection. Args: detections: Detections to generate labels for. class_names: List of class names

Get column names for the metrics that were computed. Returns all summary fields for each metric type that is available. Args: has_cl

Initialize supervision annotators based on display options. Args: show_boxes: Create BoxAnnotator. show_masks: Create MaskAnnotat

Unwrap Optional/Union/generics to base type for CLI argument parsing. Converts complex annotations like Optional[int], list[str], or int | None

Print available datasets, splits, and asset types.

Remove tracklets that have been lost for too long. Returns: List of tracklets whose time_since_update has not exceeded

Recompute Q and R from the current bbox size.

Remap original track IDs to 0-indexed contiguous values.

Resolve gt_dir and tracker_dir for MOT17 layout test data. The MOT17 test data has structure: data_path/gt/{Benchmark}-{split}/{seq}/gt/g

Resolve video output path, handling directories. If path is an existing directory, generates 'output.mp4' inside it. If path has no extension

Advance tracker until at least n confirmed tracks exist; return their IDs.

Create new tracklets from bounding boxes. Args: boxes: Bounding boxes `(N, 4)` in xyxy format.

Create new tracklets from unmatched high-confidence detections. On the very first frame, new tracklets are immediately activated with a

Apply homography transformation to points. Args: points: Array of shape `(N, 2)` containing `(x, y)` coordinates. mat

Boolean mask of ground-truth rows that are scored as ground truth. Mirrors TrackEval's ``gt_to_keep_mask``: a row counts as ground truth only

Validate output path and create parent directories if needed. Args: path: Output file path to validate. overwrite: If True, allow

Convert boxes from xywh format to xyxy format. Args: boxes: Array of shape `(N, 4)` in xywh format `(x, y, width, height)`. Returns:

Add the download subcommand to the argument parser.

Calculate the IoA (Intersection over Area) between two sets of boxes. IoA is calculated as intersection divided by the area of boxes1. This is

Compute CLEAR metrics for multi-object tracking evaluation. Calculate standard CLEAR metrics including MOTA, MOTP, ID switches, and track qual

Return the current bounding-box estimate in xyxy format.

Create a directory with *n* PNG images.

Serialize to JSON string. Args: indent: Indentation level for formatting. Defaults to `2`. Returns: JSON str

Predict the next bounding-box position. Increments ``time_since_update`` to track how many frames have elapsed since the last matched

Reset internal state. After calling reset: - The next `estimate()` call returns identity and initializes prev-frame state.

Reset tracker state by clearing all tracks and resetting ID counter. Call this method when switching to a new video or scene.

Resolve the tracker ID for the current tracklet state. Assigns a new unique ID if the tracklet is mature but hasn't been assigned one

Update tracklet state with a new bounding-box observation. Called only when the track is matched to a detection. Missed frames are ha

Factory for creating test videos with specified number of frames.

Add tracker-specific parameters from registry to argument group.

Aggregate metrics across sequences.

Return the best available PyTorch compute device, preferring acceleration. Returns: The selected device (``cuda``, ``mps``, or ``cpu``).

Build BreadcrumbList JSON-LD from navigation hierarchy. Returns None if the page is at the root level (no meaningful breadcrumb) or if the pa

Build direction consistency cost matrix for OC-SORT's OCM mechanism. Measures how well each potential (tracklet, detection) association aligns

Collect valid IDs across all frames and build original-to-0-indexed maps. Returns: Tuple of (ground_truth_id_map, tracker_id_map) where e

Instantiate the tracker from sampled and fixed parameters.

Ensure XCYCSR state keeps positive scale and aspect ratio.

Ensure XCYCWH state keeps positive width and height.

Ensure XYXY state keeps valid box corners.

Subclass hook — compute similarity for non-empty inputs. Args: boxes_1: ``(N, 4)`` array of boxes in ``[x1, y1, x2, y2]`` format.

Compute the direction consistency matrix for association, including confidence scaling.

Compute normalized direction vector between two bounding box centers. Args: bbox1: First bounding box `[x1, y1, x2, y2]`.

Set initial P, Q, R based on the first detection's size.

Configure Kalman filter noise matrices (OC-SORT paper behaviour) and SORT behaviour for XYXY coordinates.

Configure Kalman filter noise matrices (OC-SORT paper tuning).

Configure Kalman filter noise (original ByteTrack tuning).

Create and configure a Kalman filter for *bbox*. Args: bbox: First detection `[x1, y1, x2, y2]`. Returns: A

Auto-detect directory format. Args: gt_dir: Ground truth directory (direct parent of sequences). Returns: "flat" if .txt fil

Return the list of sequence names to tune over. Reads sequence names from ``seqmap`` when provided, otherwise discovers them by globbing ``*.

Discover sequence names from directory structure. Args: gt_dir: Ground truth directory (direct parent of sequences). data_format:

Boolean mask of ground-truth rows belonging to a distractor class. Mirrors TrackEval's ``distractor_classes``. Tracker detections that best-m

Download a file with a progress bar and optional MD5 verification. If `destination` already exists and `md5` is provided, the existing file's

ECC-based affine motion estimation. This method estimates a global 2D Euclidean transform between the previous frame and the

Feature affine estimation. ORB-based or SIFT-based (different initializations of self.detector, self.extractor and self.matcher for

Estimate homography transformation between point sets. Args: previous_points: Points from previous frame. current_poi

Sparse optical-flow-based affine estimation. Steps: 1) grayscale (+ optional downscale) 2) detect corners us

Extract and split ground truth data for a single frame. Returns: Tuple of (valid_boxes, valid_ids, all_boxes, distractor_mask). F

Extract tracker detections for a single frame, keeping only confirmed tracks. Returns: Tuple of (boxes, ids) with unconfirmed tracks (id

Extract tracker parameters from CLI args. Args: tracker_id: Registered tracker name. args: Parsed CLI arguments. Returns:

Extract a ZIP archive into `output_dir`.

Recursively search nav for the page, building the path of sections.

Format benchmark metrics as a table. Args: sequences: Dictionary of sequence results. aggregate: Aggregate result. column