SFSORT

Paper: SFSORT: Scene Features-based Simple Online Real-Time Tracker

SFSORT is designed around speed. The paper removes the Kalman filter entirely, introduces a bounding-box similarity cost, and uses scene-derived cues to keep association strong while minimizing compute. The goal is not to be the most elaborate tracker, but to keep the tracker extremely lightweight and real-time while still remaining competitive on standard MOT benchmarks.

Python AABB mode also supports optional EdgeTAM mask guidance with --tracker sfsort --tracker-backend python --asso-func iou --edgetam --mask-guidance-weights edgetam.pt. It requires IoU association and per_class=False, retains this tracker's existing association rules, and adds temporal model inference. Accuracy gains have not been established for this extension.

What BoxMOT Needs For SFSORT

• Detector only. ReID is not required.
• Supports both AABB and OBB detections in BoxMOT.
• Best when throughput matters most and you want a very lightweight online tracker.

Native C++ Backend

BoxMOT also ships a native C++17 SFSORT implementation under boxmot/native/cpp/trackers/sfsort/. It supports:

• cached eval and tune streamed through the live typed API
• live track through --tracker-backend cpp
• both AABB and OBB detection layouts in the native tracker path

Requirements:

• C++17 compiler
• CMake 3.16+
• OpenCV 4.x
• Eigen3 3.3+

Example:

boxmot eval --experiment mot17/ablation-yolox-lmbn.yaml --build BUILD_ID --tracker sfsort --tracker-backend cpp
boxmot track --tracker sfsort --tracker-backend cpp --source 0

Python API

Pass algorithm settings through SFSORTConfig, for example SFSORT(config=SFSORTConfig(high_th=0.6)). Direct construction and create_tracker("sfsort") use the same algorithm defaults. See the tracker configuration guide for presets and component settings.

Bases: TrackerConfig

Configure confidence stages, density adjustments, and image regions.

Shared history and association settings are inherited from TrackerConfig. Match thresholds and their dynamic adjustment apply to overlap metrics. Centroid matching uses centroid_max_distance; box cues still rank pairs.

Parameters:

• high_th (float | None, default: 0.6 ) –

  Confidence threshold for first-pass detections.

• match_th_first (float | None, default: 0.67 ) –

  Maximum geometric cost for the first association pass; smaller values require closer matches.

• new_track_th (float | None, default: 0.7 ) –

  Minimum confidence for creating a new track.

• low_th (float | None, default: 0.1 ) –

  Minimum confidence for second-pass detections.

• match_th_second (float | None, default: 0.3 ) –

  Maximum geometric cost for the second association pass.

• fuse_detection_score (bool, default: False ) –

  Weight high-confidence association similarity by detection confidence before applying mask and centroid gates.

• dynamic_tuning (bool, default: False ) –

  Adjust thresholds using the current detection count.

• cth (float | None, default: 0.5 ) –

  Confidence cutoff for detections counted by dynamic tuning.

• high_th_m (float | None, default: 0.0 ) –

  Scale for decreasing high_th during dynamic tuning.

• new_track_th_m (float | None, default: 0.0 ) –

  Scale for increasing new_track_th during dynamic tuning.

• match_th_first_m (float | None, default: 0.0 ) –

  Scale for decreasing match_th_first during dynamic tuning.

• obb_theta_damping (float, default: 0.8 ) –

  Previous angular-update weight in OBB smoothing; larger values reduce the influence of the latest angle change.

• marginal_timeout (int | None, default: 0 ) –

  Frames to retain tracks lost near an image edge.

• central_timeout (int | None, default: 0 ) –

  Frames to retain tracks lost inside the central region.

• frame_width (int | None, default: None ) –

  Optional width in pixels; configure together with frame_height, or supply frame dimensions during updates.

• frame_height (int | None, default: None ) –

  Optional height in pixels, paired with frame_width.

• horizontal_margin (int | None, default: 0 ) –

  Left and right margin size in pixels.

• vertical_margin (int | None, default: 0 ) –

  Top and bottom margin size in pixels.

__post_init__

__post_init__() -> None

Normalize default-valued inputs and validate paired dimensions.

Bases: BoxTracker

Track AABB or OBB detections with geometric matching and region timeouts.

Uses observed boxes and OBB angle smoothing without a Kalman filter. Frame dimensions define central and marginal regions; image pixels and appearance embeddings are unused.

Parameters:

• config (SFSORTConfig | None, default: None ) –

  Immutable algorithm settings. None selects SFSORTConfig defaults.

• **kwargs (Unpack[BoxTrackerOptions], default: {} ) –

  Runtime per_class, is_obb, class_ids, class_names, mask_guidance, and edgetam settings.

calculate_cost staticmethod

calculate_cost(
    tracks: list[Track],
    boxes: ndarray,
    iou_only: bool = False,
    is_obb: bool = False,
    association_function: Callable[
        [ndarray, ndarray], ndarray
    ]
    | None = None,
) -> ndarray

Calculate cost from the selected geometry and SFSORT box cues.