Training-Free Long-Term Multi-Object Tracking for Sports Video Analytics

Tomasz Stanczyk; Seongro Yoon; Francois Bremond

cs.CVAug 16, 2026

Training-Free Long-Term Multi-Object Tracking for Sports Video Analytics

Authors: Tomasz Stanczyk, Seongro Yoon, Francois Bremond

Organizations: Inria, France · Université Côte d’Azur, France

Abstract

Long-term multi-object tracking in sports remains challenging due to frequent occlusions, rapid camera motion, and repeated player reappearances. We introduce McByte++, a training-free tracking-by-detection framework that integrates lightweight mask propagation, conditional camera motion compensation, and online re-identification within a unified pipeline. Compared to its predecessor, McByte++ substantially improves runtime efficiency while enhancing identity preservation. On SoccerNet-tracking and SportsMOT benchmarks, McByte++ achieves up to +3.0 HOTA and +6.1 IDF1 improvements over the original McByte in the online setting, with further gains when combined with offline global association. Replacing heavy segmentation components and optimizing motion modeling yields up to an order-of-magnitude speed increase. All results are obtained without detector retraining or dataset-specific tuning. Code will be made available at https://github.com/tstanczyk95/McBytePlusPlus.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of the proposed McByte++ online tracking pipeline. (a) Short-term tracking. Top: Tracklet bounding boxes from the previous frame $t\!-\!1$ are treated as the tracker state and propagated to frame $t$ using a Kalman filter to generate candidate tracklet positions. Middle: Global camera motion is estimated and used to conditionally compensate the predicted tracklet positions, while the current frame $t$ is processed by the object detector. An initial association cost matrix is computed based on the IoU between compensated tracklet candidates and detection bounding boxes. Bottom: A mask temporal propagation module produces object masks at frame $t$ , which are used as an auxiliary observation to enrich the association cost matrix under regulated conditions. The final cost matrix is solved using the Hungarian matching algorithm, yielding updated tracklets at frame $t$ . (b) Long-term tracking. During short-term tracking, appearance-based re-identification features are extracted from tracked subjects using crops of their bounding boxes and stored in a feature memory. When a new tracklet is initialized at a later time $t\!+\!k$ , its appearance representation is compared against those of previously terminated tracklets. Based on feature similarity, the new tracklet can be reconnected to and extend a corresponding long-term identity.

Figure 2: (a) Examples illustrating how the bounding box coverage (

mc

) and mask fill ratio (

mf

) vary depending on the relative position of a temporally propagated mask (blue) and a detection bounding box. The most favorable case for mask-based guidance occurs when both

mc

and

mf

are close to

1

. (b) Illustration of ambiguity and isolation cases in the association process, where temporally propagated masks can provide additional guidance. Ambiguity arises when IoU-based costs are similarly low for multiple entries in a row or column of the cost matrix, while isolation occurs when relevant IoU-based costs are too high to allow association despite the absence of ambiguity.

Unconditional camera motion compensation – no re-ID
Method	HOTA $\uparrow$	IDF1 $\uparrow$	MOTA $\uparrow$	FPS $\uparrow$
McByte [ 20 ]	85.0	79.9	96.8	1.04
McByte++, ds=2 $\dagger$	84.6	79.5	97.0	6.49
McByte++, ds=4	84.1	78.9	97.1	10.51
McByte++, ds=6	83.4	78.0	97.0	10.33
Conditional camera motion compensation – no re-ID

Table 1: Ablation study on SoccerNet-tracking 2022 (test split) evaluating the impact of mask replacement, camera motion compensation (CMC) strategy, spatial downscaling factor (

ds

), and online re-identification. Results are reported using identical oracle detections. Runtime (FPS) improvements resulting from lightweight mask propagation and optimized CMC are explicitly reported. The original McByte is highlighted in yellow. The selected default McByte++ configuration (conditional CMC with

ds=4

) is highlighted in green. The best online re-ID configuration and the offline re-ID configuration are highlighted in blue.

Unconditional camera motion compensation – no re-ID
Method	HOTA $\uparrow$	IDF1 $\uparrow$	MOTA $\uparrow$	FPS $\uparrow$
McByte [ 20 ]	76.9	77.5	97.2	3.60
McByte++, ds=2 $\dagger$	75.9	76.3	96.9	11.88
McByte++, ds=4	75.9	76.3	96.9	18.65
McByte++, ds=6	72.2	71.3	96.2	13.68
Conditional camera motion compensation – no re-ID

Table 2: Ablation study on SportsMOT (test split) under identical detection settings. The table evaluates unconditional and conditional CMC, spatial downscaling factors (

ds

), and online re-identification thresholds. Runtime improvements resulting from lightweight mask propagation and optimized CMC are explicitly reported. Color highlights follow the same convention as in Table 1 .

Online processing
Method	HOTA $\uparrow$	IDF1 $\uparrow$	MOTA $\uparrow$	FPS $\uparrow$
McByte [ 20 ]	64.1	76.5	81.8	1.46
McByte++, no re-ID	62.4	74.1	81.7	15.05
McByte++, with re-ID	64.3	78.6	81.8	11.13
Offline, with postprocessing
McByte++ with GTA	65.7	80.7	81.7	10.06

Table 3: Results on SoccerNet-tracking Challenge 2023 split using YOLOX detections trained on SportsMOT. Due to unexpected unavailability of the official evaluation server during preparation of this work, comparisons are limited to McByte variants. McByte++ demonstrates improved identity preservation and substantially higher runtime efficiency compared to the original McByte.

Panel of Figure 3 — Figure 3: Qualitative example of long-term identity preservation on SoccerNet-tracking. Players temporarily leave the field of view and reappear after several frames. McByte++ successfully reconnects the tracklets using online re-identification, preserving the original identity without introducing an ID switch. Bounding boxes are shown for clarity, while masks predicted by EdgeTAM are visualized only to illustrate spatial alignment and how they are used by McByte++. This example highlights the effectiveness of the online long-term identity association mechanism in realistic sports scenarios with frequent occlusions and camera motion.

Method	Train	HOTA $\uparrow$	IDF1 $\uparrow$	MOTA $\uparrow$
The same detections
ByteTrack [ 11 ]	✕	72.1	75.3	94.5
OC-SORT [ 12 ]	✕	82.0	76.3	98.3
McByte [ 20 ]	✕	85.0	79.9	96.8
McByte++, online	✕	87.5	84.5	97.1
McByte++, offline with GTA	✕	88.6	87.2	97.1

Table 4: Comparison with state-of-the-art tracking methods on SoccerNet-tracking 2022 (test split). Oracle detections are used when provided. McByte++ operates without any detector retraining or dataset-specific tuning. HOTA and IDF1 are emphasized as primary identity and association metrics. The best results per metrics are in bold and second-best are underlined.

Method	Train	HOTA $\uparrow$	IDF1 $\uparrow$	MOTA $\uparrow$
Detections on their own
GTR [ 52 ]	✓	54.5	55.8	67.9
CenterTrack [ 53 ]	✓	62.7	60.0	90.8
MeMOTR [ 5 ]	✓	70.0	71.4	91.5
MOTIP [ 7 ]	✓	71.9	75.0	92.9
MotionTrack [ 54 ]	✓	74.0	74.0	96.6

Table 5: Comparison with state-of-the-art tracking methods on SportsMOT (test split). Results are reported under identical detection settings when possible. McByte++ achieves the highest HOTA and IDF1 among multi-object tracking approaches while maintaining a fully training-free pipeline. The best results per metrics are in bold and second-best are underlined.

Method	HOTA $\uparrow$	IDF1 $\uparrow$	MOTA $\uparrow$
DEVA, original settings	39.3	37.3	-109.6
DEVA, with YOLOX	42.4	42.1	-57.0
Grounded SAM 2, original settings	45.9	44.7	-13.9
Grounded SAM 2, with YOLOX	66.1	70.2	91.4
MASA, original settings	39.4	35.7	-27.2
MASA, with YOLOX	73.6	71.2	97.0

Table 6: Comparison with segmentation-based tracking systems on SportsMOT validation set . Results are reported using both original method settings and YOLOX detections trained on SportsMOT. McByte++ integrates lightweight mask propagation within a structured tracking-by-detection framework and achieves superior identity preservation compared to pure mask-based approaches. The original McByte is also listed for reference. The best results per metrics are in bold and second-best are underlined.

Multi-Object Tracking

Multi-object tracking (MOT) links detections across video frames to maintain object identities through motion, occlusion, and changing viewpoints.

See all 189 papers

SAMIDARE: Advanced Tracking-by-Segmentation for Dense Scenarios
Apr 24, 2026Shozaburo Hirano, Norimichi UkitaSports Analytics Visual Object Tracking
Toyota Technological Institute
ARC-SORT: Adaptive, Robust, Compensated Multi-Object Tracking with Interacting Multiple Models
Oct 15, 2026Girish Kumar AdariMulti-Object Tracking
Athlitix Inc.
ByteTraX: Enhancing the ByteTrack Architecture with Optimised Thresholding
Sep 29, 2026Thomas A. O'Shea-WhellerData Association Multi-Object Tracking
University of Exeter

Apr 24, 2026J/K move · Enter open · S save

SAMIDARE: Advanced Tracking-by-Segmentation for Dense Scenarios

Shozaburo Hirano, Norimichi Ukita

Toyota Technological Institute

Sports Analytics Visual Object Tracking Sports Video Analysis Segment Anything Models Multi-Object Tracking

Open paper AskPDF arXiv

Automated sports analysis demands robust multi-object tracking (MOT), yet segmentation-based methods often struggle with mask errors and ID switches in dense scenes. We propose SAMIDARE, a framework that enhances SAM2MOT for crowded scenes through three key components: (1) density-aware mask re-generation and (2) selective memory updates, both for adaptive mask control to preserve target feature integrity, and (3) state-aware association and new track initialization, which improves robustness under mutual occlusions and frequent frame-out events. Evaluated on the SportsMOT dataset, SAMIDARE achieves state-of-the-art performance, outperforming the baseline by 2.5 HOTA and 4.2 IDF1 points on the validation set. These results demonstrate that adaptive feature management using mask control and state-aware association provide a robust and efficient solution for dense sports tracking. Code is available at https://github.com/ZabuZabuZabu/SAMIDARE

Abstract

Multi-Object Tracking

Explore similar work

SAMIDARE: Advanced Tracking-by-Segmentation for Dense Scenarios

ARC-SORT: Adaptive, Robust, Compensated Multi-Object Tracking with Interacting Multiple Models

ByteTraX: Enhancing the ByteTrack Architecture with Optimised Thresholding