Skip to content

CLI Reference

Every command the trackers CLI exposes, in one table, plus the full flag reference for trackers inspect's four subcommands — the one part of the CLI with no flag table anywhere else in the docs.

What you'll learn:

  • Every top-level command and subcommand the CLI exposes
  • Where each command's detailed guide and flag table lives
  • The full flag reference for trackers inspect sam / cutie / mask-manager / mcbyte

Install

Get started by installing the package.

pip install trackers

inspect subcommands additionally need the mask extra — see Requires the mask extra below. For more options, see the install guide.


Commands

Command Purpose Reference
trackers track Run a tracker on a video, webcam, RTSP stream, or image directory. Track guide
trackers eval Score tracker output against ground truth with CLEAR, HOTA, and Identity metrics. Evaluate guide
trackers download Download MOT benchmark datasets — MOT17 and SportsMOT. Download guide
trackers tune Search tracker hyperparameters with Optuna. Tune guide
trackers inspect sam Run SAM mask generation on a single image and save a visualization. Below
trackers inspect cutie Seed masks with SAM on the first frame, then propagate them with Cutie across a frame range. Below
trackers inspect mask-manager Run the real MaskManager orchestration — initialize, propagate, add, remove — in manual or gt mode. Below
trackers inspect mcbyte Compare locked-IoU vs. mask-conditioned McByte on one sequence. Below
trackers benchmark mcbyte Run McByte over complete benchmark test sets and write MOTChallenge-format results. Benchmark Runner guide

For a narrative walkthrough of track, eval, download, or tune, see their own guides linked above. For inspect, see the Inspect the Mask Pipeline guide for worked examples — the flag tables below cover the full surface of each subcommand.

Every trackers subcommand accepts --config <file> to load arguments from a YAML file, and hyphens and underscores are interchangeable in flag names (--cmc-downscale and --cmc_downscale are the same option). Boolean flags gain a negated --no_<name> counterpart. A nested option group (e.g. --tracker.mask.*) exposes a .help sub-command — --tracker.mask.help lists that group's fields without printing the rest of --help.


Requires the mask extra

All four inspect subcommands build SAM and Cutie, so install them first:

pip install "trackers[mask]"

Without the extra, the command exits with an install hint rather than a traceback. For more options, see the install guide.


Inspect Subcommands

Each tab lists all arguments accepted by that trackers inspect subcommand.

trackers inspect sam

Argument Description Default
--image_path Path to the input image.
--box Bounding boxes in xyxy format, given as one list: --box='[[x1,y1,x2,y2]]'. Append with --box+=....
--output_path Path to save the visualization. outputs/inspect/sam/sam_masks.jpg
--device Device used by SAM, e.g. cpu or cuda. auto resolves to CUDA when available, otherwise CPU. auto
--model_type SAM model type. vit_b

trackers inspect cutie

Argument Description Default
--image_dir Directory containing input frames.
--start_file First frame filename, included in the selected frame range.
--end_file Last frame filename, included in the selected frame range.
--box Bounding boxes on the first selected frame, xyxy, as one list. Append with --box+=....
--add_at Masks to add via a box on a given frame, each filename:x1,y1,x2,y2. Applied on that frame then propagated to the next, matching McByte timing; each event is treated as a new object. none
--remove_at Masks to remove before propagating to the given frame, each filename:manual_mask_id. none
--output_root Root directory for timestamped outputs. outputs/inspect/cutie
--device Device used by SAM and Cutie, e.g. cpu/cuda. auto resolves to CUDA when available, otherwise CPU. auto
--sam_model_type SAM model type. vit_b
--cutie_model_type Cutie model type. base-mega
--cutie_config_path Optional path to Cutie's Hydra config directory. none
--cutie_config_name Cutie Hydra config name. eval_config

trackers inspect mask-manager--mode selects manual or gt; passing an option that belongs to the other mode is an error.

Argument Description Default
--image_dir Directory holding the frame images.
--mode manual (boxes and lifecycle events from the CLI) or gt (replay from a MOT-format ground-truth file).
--start_file Manual mode. Filename of the first frame to process. none
--end_file Manual mode. Filename of the last frame to process. none
--box Manual mode. Initial tracklet boxes on the first selected frame, xyxy, as one list. Append with --box+=.... none
--add_at Manual mode. New tracklets to add from given frame boxes, each filename:x1,y1,x2,y2. none
--remove_at Manual mode. Tracklets to remove before propagating to the given frame, each filename:tracklet_id. none
--gt_file GT mode. MOT-format ground-truth file to replay. none
--start_frame GT mode. First frame number to replay. none
--end_frame GT mode. Last frame number to replay. none
--tracklet_id GT mode. Tracklet IDs to replay, e.g. --tracklet_id='[3,7]'. Omit, or pass [all], to replay every tracklet. all
--output_root Directory holding run directories, grouped by mode (<output_root>/<mode>/<timestamp>/). outputs/inspect/mask-manager
--device Device used by SAM and Cutie, e.g. cuda/cpu. auto resolves to CUDA when available, otherwise CPU. auto
--sam_model_type SAM model type. vit_b
--cutie_model_type Cutie model type. base-mega
--cutie_config_path Directory holding the Cutie config. Omit to use the config shipped with the installed Cutie package. none
--cutie_config_name Name of the Cutie config to load. eval_config
--mask_creation_bbox_overlap_threshold GT mode. Overlap threshold above which mask creation is delayed. 0.6

trackers inspect mcbyte — options are grouped into nested dataclasses: --sequence.*, --cmc.*, --mask.*.

Argument Description Default
--sequence.image_dir Directory containing sequence frames.
--sequence.det_file Path to the detection file.
--sequence.start_frame First frame number to process, inclusive.
--sequence.end_frame Last frame number to process, inclusive.
--sequence.det_format Detection-file column format: mot_tlwh (frame,id,left,top,width,height,confidence,...) or xyxy (frame,x1,y1,x2,y2,confidence). mot_tlwh
--sequence.frame_rate Sequence frame rate used to scale the lost-track buffer. 30.0
--cmc.enable Enable camera motion compensation in both runs. false
--cmc.method Camera-motion compensation method. Options: orb, sift, sparseOptFlow, ecc. sparseOptFlow
--cmc.downscale Image downscale factor used by CMC. 6
--mask.device Device used by SAM and Cutie in the mask-conditioned run. auto resolves to CUDA when available, otherwise CPU. auto
--mask.enable_isolated_matching Allow mask evidence to rescue isolated positive-IoU pairs below the normal association threshold. false
--modes Tracker configurations to run: --modes=[locked_iou]. both locked_iou and mask_conditioned
--output_dir Directory holding one timestamped run directory per invocation; both comparison runs write inside it. outputs/inspect/mcbyte