PaperlayerView exact revision on arXiv

Source-linked AI summary

ByteTraX: Enhancing the ByteTrack Architecture with Optimised Thresholding

Thomas A. O'Shea-Wheller

arXiv:2609.37801v2cs.CV

TL;DR

Tracking-by-detection algorithms remain broadly adopted for reduced computational requirements and ease of integration, making their optimization a practical route to improve established deployment pipelines. ByteTraX relaxes matching thresholds and, in closed systems, can use track reconnection and merging to preserve track continuity. Across closed-system benchmarks, ByteTraX increases HOTA by 5.653 and IDF1 by 7.987, while reducing mean IDSW from 103 to 53 and increasing FPS by 7.044 versus ByteTrack.

  • Problem

    Tracking-by-detection algorithms remain broadly adopted for reduced computational requirements and ease of integration, making their optimization a practical route to improve established deployment pipelines.

    Source: §1 · p. 1

  • Method

    ByteTraX relaxes matching thresholds and, in closed systems, can use track reconnection and merging to preserve track continuity.

    Source: §1 · p. 2, §3.1 · pp. 4–5

  • Results

    Across closed-system benchmarks, ByteTraX increases HOTA by 5.653 and IDF1 by 7.987, while reducing mean IDSW from 103 to 53 and increasing FPS by 7.044 versus ByteTrack.

    Source: §4.4 · pp. 8–9

  • Takeaways & Limitations

    ByteTraX is presented as an improvement to the widely used ByteTrack baseline, with an integration pipeline for YOLO26 and source code for expansion to other detection frameworks.

    Source: §5 · p. 10

    The evaluation compares ByteTraX only with ByteTrack, using the same upstream YOLO26 models, rather than testing against other tracking architectures.

    Source: §4.4 · p. 8

Abstract

from arXiv · show

The ByteTrack algorithm is a widely used and computationally efficient multi-object tracking architecture. Its core innovation lies in the combination of lenient bounding box associations with tracklet similarity matching to robustly deal with object occlusions. However, this strategy is nevertheless vulnerable to erroneous track reclassification and identity switching, as detection confidence scores dictate association priority. To address this, I present a simple enhancement of the ByteTrack architecture--named ByteTraX--that optimises track continuity via a single unified matching threshold, while penalising identity switches through stringent track initiation criteria. This approach achieves consistently improved performance across a range of diverse benchmarks including GMOT-40, LC-MOT, SportsMOT, TeamTrack, DAMUNT, and DeepSea-MOT, while simultaneously increasing processing speed by >10%. Specifically, results demonstrate a >40% reduction in identity switches, accompanied by mean increases in HOTA of 3.6, IDF1 of 5.6, and FPS of 6.3. As such, adoption of the ByteTraX algorithm has the potential to substantially enhance tracking performance over the ByteTrack baseline, while retaining the efficiency needed for real-time deployment. To facilitate usage, I provide the source code, integration functionality for the YOLO family of object detection models, and deployment instructions via an open source repository.

1. Introduction

Tracking-by-detection remains widely adopted for its efficiency and ease of integration, but ByteTrack can lose or reassign identities when confidence-based thresholding eliminates low-confidence detections. ByteTraX reduces confidence’s role in matching and adapts its strategies to open and closed systems, with reported improvements in performance and speed across six benchmarks.

1. Introduction

  • ByteTrack’s mistaken elimination or supersedure of low-confidence detections can cause track reclassification and identity switches, undermining long-term identity preservation.The problem is amplified as track length and motion complexity increase; low-confidence detections can still represent true object positions.Source: §1 · pp. 1–2
  • Because confidence scores alone cannot reliably distinguish true detections from false positives across models and deployment contexts, the introduction motivates using additional tracking cues.True and false detections can have overlapping confidence distributions, and filtering true low-confidence detections can cause irreversible tracking errors.Source: §1 · p. 2
  • ByteTraX treats detections equally, uses permissive matching thresholds, predicts with broad uncertainty bounds, and applies bespoke strategies across open and closed systems.The comparison uses the same YOLO26 front end for both trackers and evaluates six benchmarks with HOTA, CLEAR, ID, and line-crossing measures.Source: §1 · pp. 2–3
  • ByteTraX is reported to improve performance and speed over ByteTrack across benchmarks spanning diverse tracking scenarios.The benchmarks include GMOT-40, LC-MOT, SportsMOT, TeamTrack, DAMUNT, and DeepSea-MOT.Source: §1 · p. 3

2. Related Work

Tracking-by-detection links detections across frames using motion prediction, similarity scoring, and matching, but confidence thresholds can discard valid detections. ByteTrack addresses this with two-stage matching, while ByteTraX targets the remaining identity-reassignment problem by reducing confidence’s influence on associations.

2.1. Tracking-by-detection

  • Tracking-by-detection predicts track locations, scores candidate detections by overlap or appearance, and matches candidates to tracks across frames.Its efficiency and integration flexibility support its use in real-time tracking pipelines.Source: §2.1 · p. 3
  • Occlusion and changing detection confidence can cause correct low-scoring detections to be discarded before they enter the tracking process.The difficulty arises because confidence changes may be independent of whether an object is present.Source: §2.1 · pp. 3–4

2.2. ByteTrack

  • ByteTrack’s BYTE association matches high-confidence detections first, then uses remaining low-confidence detections to associate unmatched tracklets.This two-stage procedure recovers partially occluded detections that high-score-only methods would miss.Source: §2.2 · p. 4
  • ByteTrack still prioritizes matches by detection confidence, which the paper links to track breaks and subsequent ID reassignment when confidence fluctuates.ByteTraX therefore reduces confidence’s contribution to matching and association thresholds to prioritize identity preservation and track continuity.Source: §2.2 · p. 4

3. Methods

ByteTraX relaxes association thresholds and gives existing tracks greater weight, while optional closed-system heuristics reconnect lost tracks and merge duplicate tracklets. The heuristics are designed to supplement standard association when it cannot reliably resolve a track.

3.1. ByteTraX Architecture

  • ByteTraX removes ByteTrack’s staged confidence thresholds and lowers the IoU similarity requirement, allowing detections near predicted tracks to associate despite uncertain motion predictions.The combined changes favor existing tracks and reduce new ID creation; closed-system reconnection and merging functions can further support recovery.Source: §3.1 · pp. 4–5

3.2. Track Reconnection

  • In closed systems, reconnection links a lost tracklet to a same-class detection within time and distance thresholds, even when motion projections fail.The heuristic activates after standard association fails; for multiple candidates, matching is filtered by proximity to the last known location and can address extended occlusion or detection failure.Source: §3.2 · p. 5

3.3. Tracklet Merging

  • Tracklet merging screens new tracks in fixed-count systems and merges qualifying same-class tracks that overlap existing trajectories.It runs after standard association and reconnection, limiting merges to detections those steps could not reliably assign; the heuristics can be enabled optionally for closed systems.Source: §3.3 · p. 5

4. Experiments

ByteTraX is evaluated against ByteTrack across six open- and closed-system benchmarks using tracking, identity, speed, and counting measures. It improves reported tracking results in both system types; closed-system track reconnection and merging improve them further.

4.1. Datasets

  • The six benchmarks cover open systems, where tracks enter or leave view, and closed systems, where all tracks remain visible, across diverse scenarios and motion patterns.Sequences range from 4 to 100 co-occurring tracks, with mean speeds of 0.2–89 pixels per second and occlusion rates of 0.010–0.983.Source: §4.1 · p. 6

4.2. Metrics

  • Evaluation combines HOTA, MOTA, IDF1, IDSW, excess ID count, and FPS with a line-crossing measure for a common counting-by-tracking task.The measures cover tracking accuracy, identity preservation, computational efficiency, and trajectory-crossing counts against ground truth.Source: §4.2 · pp. 6–7

4.3. Implementation

  • Both trackers use benchmark-specific YOLO26n detectors trained for 100 epochs with a common 40:10:50 split; only ByteTraX-RT adds reconnection and merging on closed-system datasets.Comparisons focus on ByteTrack and ByteTraX, using the same upstream detector models for parity.Source: §4.3 · pp. 7–8, §4.4 · p. 8

4.4. Comparison with ByteTrack

  • On closed-system benchmarks, ByteTraX’s HOTA score increases by 5.653 over ByteTrack; IDF1 increases by 7.987 and MOTA by 0.968.Table 2 values are combined means averaged over all benchmark datasets and detection confidence thresholds belonging to each system type.Source: §4.4 · p. 8, Table 2 · p. 8
  • Open-system ByteTraX raises IDF1 by 2.575 and FPS by 5.170, while mean IDSW falls from 91 to 58 and excess IDs from 1013 to 447.Aggregate line-crossing accuracy also increases by 1.768% over ByteTrack.Source: Figure 3 · p. 7, §4.4 · p. 8
  • ByteTraX gains vary by dataset and threshold: DeepSea-MOT shows no meaningful MOTA gain or IDSW reduction, with IDSW falling only at thresholds 0.50 and 0.75.ByteTraX-RT’s benefits over standard ByteTraX are comparatively diminished at 0.75; gains are greatest for DAMUNT and TeamTrack and more modest for LC-MOT.Source: §4.4 · pp. 8–9
  • Closed-system ByteTraX raises IDF1 by 7.987 and FPS by 7.044, with mean IDSW falling from 103 to 53 and excess IDs from 137 to 109.Mean line-crossing accuracy increases by 7.530%.Source: §4.4 · pp. 8–9, Figure 4 · p. 9
  • ByteTraX-RT improves on standard ByteTraX across all metrics, including HOTA by 3.102 and IDF1 by 4.224, while mean IDSW falls from 53 to 30.Its excess ID count falls to 36, and its mean line-crossing accuracy is 3.189% higher than standard ByteTraX.Source: §4.4 · p. 9, Figure 4 · p. 9

5. Conclusion

ByteTraX reports improved tracking performance and speed across diverse scenarios, aiming to improve on ByteTrack rather than challenge state-of-the-art methods. The paper provides a YOLO26 integration pipeline and source code for expansion to other detection frameworks.

5. Conclusion

  • ByteTraX delivers substantially improved performance and speed across diverse tracking scenarios.The conclusion spans applications from insect-colony trajectory resolution to deep-ocean species quantification.Source: §5 · pp. 9–10
  • ByteTraX aims to improve the widely used ByteTrack baseline, not to challenge the state of the art.Source: §5 · p. 10
  • The release includes a full YOLO26 integration pipeline and source code for expansion to additional detection frameworks.Source: §5 · p. 10

Related papers

Loading 2609.37801v2…