ByteTraX: Enhancing the ByteTrack Architecture with Optimised Thresholding

Thomas A. O'Shea-Wheller

cs.CVSep 29, 2026

ByteTraX: Enhancing the ByteTrack Architecture with Optimised Thresholding

Authors: Thomas A. O'Shea-Wheller

Organizations: University of Exeter

Abstract

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.

Figures & tables

Figure 1. Comparison of tracking ID losses across example video sequences for (a) ByteTrack, and (b) ByteTraX. The top two panels display a sequence from LC-MOT, a closed system dataset with fixed track counts; while the bottom two panels are from GMOT-40, an open system benchmark in which tracks can enter and leave the frame. Lines indicate tracked trajectories, coloured by ID, and white arrows denote instances of ID loss. ID losses occur when bounding box detection scores fall below a tracking algorithm’s matching or similarity threshold, which can lead to the rapid accumulation of identity errors over time.

Figure 2. Example video sequences from the (a) open, and (b) closed system benchmarks used in performance evaluations. The top three panels detail sequences from SportsMOT, DeepSea-MOT, and GMOT-40; while the bottom three display sequences from LC-MOT, DAMUNT, and TeamTrack. Bounding boxes indicate ground truth tracks, coloured by ID, and green lines denote the crossing counter regions utilised in the calculation of line crossing accuracy.

Dataset	Video Length	Track Count	Track Density (Tracks/ MP)	Track Speed (Px/ Frame)	Track Angular Acceleration (°/ Frame)	Track Occlusion Rate	Track Normalised Innovation Residual	Target Size (Px)	Class Count
DeepSea-MOT [ 21 ]	599	35	7.183	4.170	36.170	0.268	0.320	51542.470	37
DAMUNT [ 20 ]	573	23	40.453	1.810	59.360	0.677	0.160	2019.370	1
GMOT-40 [ 16 ]	237	49	14.830	10.040	21.180	0.561	0.740	12989.050	1
LC-MOT [ 17 ]	1124	15	11.398	1.030	55.200	0.400	0.380	4634.940	1
SportsMOT [ 18 ]	575	14	12.004	6.070	27.000	0.593	0.390	6756.340	1
TeamTrack [ 19 ]	900	23	2.773	1.880	20.540	0.126	0.160	922.680	1

Table 1. Comparison of attributes across benchmark datasets. Values represent means averaged across all video sequences within each benchmark. Metrics are selected to outline target size, density, speed, propensity for sudden changes in acceleration and direction, occlusion rate, divergence from linear motion model predictions, and simultaneous class cooccurrence.

Figure 3. (a) Comparison of tracking performance across detection confidence thresholds for open system benchmarks ( N =53). Each point represents the combined mean score for GMOT-40, SportsMOT, and DeepSea-MOT, coloured by tracker (ByteTrack, grey; ByteTraX, light blue). (b) Comparison of tracking performance across detection confidence thresholds for closed system benchmarks ( N =15). Each point represents the mean score averaged across LC-MOT, TeamTrack, and DAMUNT, coloured by tracker (ByteTrack, grey; ByteTraX, light blue; ByteTraX-RT, dark blue).

System Type	Tracker	HOTA	MOTA	IDF1	Crossing Accuracy	IDSW	FPS
Open	ByteTrack	62.593	73.634	72.294	80.899	91	44.328
	ByteTraX	64.143	75.002	74.869	82.667	58	49.498
Closed	ByteTrack	68.356	83.629	69.832	76.379	103	51.601
	ByteTraX	74.009	84.597	77.819	83.909	53	58.645
	ByteTraX-RT	77.111	86.185	82.043	87.098	30	59.265

Table 2. Comparison of tracking performance across open and closed systems. Values represent combined means averaged over all benchmark datasets and detection confidence thresholds belonging to each system type, with the highest scoring results shown in bold. The overall tracking algorithm rankings are determined based upon the proportion of top scoring metrics achieved by each.

Figure 4. (a) Change in IDSW and excess ID counts relative to ByteTrack across detection confidence thresholds for open system benchmarks ( N =53). Boxplots indicate the median percentage change in proportion to the corresponding value for ByteTrack, averaged across GMOT-40, SportsMOT, and DeepSea-MOT, and coloured by tracker (ByteTraX, light blue) and metric (IDSW, opaque; ID Excess, transparent). (b) Change in IDSW and excess ID counts relative to ByteTrack across detection confidence thresholds for closed system benchmarks ( N =15). Boxplots again indicate the median percentage change in proportion to the corresponding value for ByteTrack, averaged across LC-MOT, TeamTrack, and DAMUNT, and coloured by tracker (ByteTraX, light blue; ByteTraX-RT, dark blue) and metric (IDSW, opaque; ID Excess, transparent). Error bars represent 1.5 times the interquartile range from the median. Positive values indicate an increase in the given metric, negative values a decrease, and a score of 0 absolute parity with ByteTrack. ID excess is defined as the number of additional IDs generated by a tracker in comparison to ground truth, while IDSW is a count of all instances in which the same ground truth tracks are assigned sequentially different IDs.

System Type	Dataset	Tracker	HOTA	MOTA	IDF1	Crossing Accuracy	IDSW	FPS
Open	GMOT-40 [ 16 ]	ByteTrack	62.557	77.971	74.519	82.880	107	46.476
		ByteTraX	64.379	79.775	77.562	84.577	65	51.720
	SportsMOT [ 18 ]	ByteTrack	62.049	90.146	71.400	78.559	67	57.249
		ByteTraX	63.304	90.930	74.279	82.086	44	66.475
	DeepSea-MOT [ 21 ]	ByteTrack	62.898	51.331	63.661	73.679	37	31.857
		ByteTraX	63.452	51.128	64.272	75.203	38	35.517

Table 3. Comparison of tracking performance across benchmark datasets. Values represent combined means averaged over all detection confidence thresholds, with the highest scoring results for each dataset shown in bold. The overall tracking algorithm rankings are determined based upon the proportion of top scoring metrics belonging to each.

Explore similar work

When Fish Look Alike: Tracking Identities with Dual-branch Elasticity
Jul 29, 2026Vran Lee, Xin Liu, Yijie Wei +3Multi-Object Tracking Object Detection
China Agricultural University · Beijing Normal University · National University of Singapore
Selective Mask Propagation for Multi-Object Tracking
Jun 11, 2026Alexander HolmbergMulti-Object Tracking Video Object Segmentation
KTH Royal Institute of Technology
Seg2Track++: Probabilistic Track Validation and Data Association for Multi-Object Tracking and Segmentation
Jun 2, 2026Diogo Mendonça, Tiago Barros, Cristiano Premebida +1Multi-Object Tracking Object Detection
University of Coimbra, Institute of Systems and Robotics, Department of Electrical and Computer Engineering, Portugal.

Jul 29, 2026, cs.CVJ/K move · Enter open · S save

When Fish Look Alike: Tracking Identities with Dual-branch Elasticity

Vran Lee, Xin Liu, Yijie Wei, Yeqiang Liu+2

China Agricultural University · Beijing Normal University · National University of Singapore

Multi-Object Tracking Object Detection Fish

Open paper AskPDF arXiv

Tracking dense, homogeneous targets like schooling fish remains a major challenge for multiple object tracking due to extreme inter-individual homogeneity, severe physical clustering, and rapid non-rigid deformations. While heavy-backbone separated detection and embedding trackers like SU-T push accuracy boundaries using complex Re-Identification networks, their computational overhead prohibits edge deployment. Furthermore, these modules often fail when appearance features degrade under severe occlusions. To overcome this, we propose Tracking Identities with Dual-branch Elasticity (TIDE). Bypassing expensive appearance cues, TIDE utilizes the Adaptive Geometric Correspondence IoU, an association mechanism leveraging spatial and structural consistency to robustly handle complex morphological variations. Crucially, TIDE introduces system-level deployment elasticity, decoupling the algorithmic pipeline from strict hardware constraints. Evaluations on the MFT-Edge benchmark demonstrate that our Lightweight L-branch achieves a competitive HOTA of 28.43 using merely 20.47G FLOPs. This represents a 38.7-fold computational reduction compared to upper bounds like SU-T, directly facilitating real-time edge deployment. Concurrently, our Scalable S-branch establishes a 29.98 HOTA, successfully bridging the gap between high-precision cloud analysis and efficient edge tracking. The dataset and codes are released at https://vranlee.github.io/TIDE/.

Abstract

Explore similar work

When Fish Look Alike: Tracking Identities with Dual-branch Elasticity

Selective Mask Propagation for Multi-Object Tracking

Seg2Track++: Probabilistic Track Validation and Data Association for Multi-Object Tracking and Segmentation