GTA-Link: Global Tracklet Association for MOT

• GTA-Link is a global association framework that refines multi-object tracking by merging fragmented tracklets and splitting those with identity mix-ups.
• It leverages appearance embeddings, spatial constraints, and association graph modeling to improve tracklet connectivity and reduce identity switches.
• Empirical results demonstrate notable improvements in HOTA scores and reduced ID switches on benchmarks like SportsMOT and SoccerNet.

Global Tracklet Association (GTA-Link) is a suite of global association methods designed to refine multi-object tracking (MOT) outputs by merging tracklets corresponding to the same identity and splitting those contaminated by identity switches. Originally introduced for sports player and automotive multi-object tracking, GTA-Link corrects both short-term fragmentation due to occlusion and long-term identity drift by leveraging appearance embeddings—with optional integration of spatial constraints and advanced association graph modeling. GTA-Link serves as a post-processing, plug-and-play module, demonstrated to improve identity consistency, reduce ID switches, and increase higher-order association metrics such as HOTA in a range of academic benchmarks (Sun et al., 2024, Jian et al., 31 Jan 2026, Nguyen et al., 2022).

1. Problem Definition and Objectives

Global Tracklet Association addresses the aggregation of fragmented, short tracklets—produced by base detection and online association algorithms—into temporally extensive, identity-coherent trajectories. For a sequence of $T$ video frames and a set of $M$ initial tracklets, each defined as

$T_i = \{(b^i_t, f^i_t) : t \in \Pi_i\}$

with $b^i_t \in \mathbb{R}^4$ the bounding box and $f^i_t \in \mathbb{R}^d$ the associated appearance embedding, the goal is to globally assign unique identity labels so that all detections corresponding to the same entity are grouped, and erroneous concatenations are resolved. The problem is often formalized as a combinatorial optimization over a binary association matrix $X \in \{0,1\}^{M \times M}$: $\min_{X}\; \sum_{i < j} D_{ij} X_{ij}$ subject to $X_{ij} \in \{0,1\}$, $X_{ii} = 1$, and transitivity (i.e., if $X_{ij} = X_{jk} = 1$, then $M$0). This formulation penalizes associations according to a tracklet-level dissimilarity $M$1 (Sun et al., 2024).

2. Appearance and Similarity Modeling

Central to GTA-Link is the use of appearance features extracted from each detection. For sports scenarios, an OSNet-based ReID model produces a per-box descriptor $M$2. The inter-tracklet affinity is computed as the mean pairwise cosine distance: $M$3 A high similarity score $M$4 indicates a likely match. In multi-camera tracking (Nguyen et al., 2022), node features are enhanced by transformers employing cross-attention between tracklets and new detections, and appearance ReID losses are imposed to structure the embedding space.

Spatial constraints further augment the affinity model in some variants, either via spatial gating on entry/exit positions (Sun et al., 2024) or by using spatial overlap metrics such as Expansion-IoU (EIoU) in GTATrack (Jian et al., 31 Jan 2026).

3. Global Tracklet Association Algorithms

The GTA-Link pipeline generally follows a two-stage algorithmic structure:

3.1 Tracklet Splitter

The splitter corrects "mix-up" errors, in which a tracklet contains detections from multiple identities. By clustering the internal appearance descriptors of each tracklet via DBSCAN (density-based spatial clustering) or other clustering strategies, inconsistent segments are detected and the tracklet is split accordingly. Parameters—such as minimum samples, neighborhood radius (cosine distance $M$5), and a cap on cluster number—control the fragmentation (Sun et al., 2024, Jian et al., 31 Jan 2026).

3.2 Tracklet Connector

The connector solves the "cut-off" problem, merging spatially- and appearance-similar fragments. For all candidate fragments, a pairwise distance matrix $M$6 is built. Merges are forbidden for fragments overlapping in time. Hierarchical single-linkage clustering or DBSCAN is used to group fragments, with merging thresholds on appearance ($M$7) and spatial gates ($M$8) preventing erroneous associations. Agglomerative merging subject to spatial, temporal, and (optionally) EIoU constraints forms full-length identity chains (Sun et al., 2024, Jian et al., 31 Jan 2026).

A common workflow is thus:

$T_i = \{(b^i_t, f^i_t) : t \in \Pi_i\}$0

4. Applications and Integration

GTA-Link is implemented as a tracker-agnostic post-processing module, requiring only a set of initial tracklets with per-detection embeddings. It integrates seamlessly with MOT baselines such as SORT, ByteTrack, and Deep-EIoU. In sports tracking pipelines like GTATrack (Jian et al., 31 Jan 2026), the output of a motion-agnostic online tracker serves as input to GTA-Link, enforcing both short-term matching and long-term identity consistency. In autonomous vehicle scenarios (Nguyen et al., 2022), the algorithm leverages multi-camera and 3D detection information, applying global association via graph-structured link prediction using cross-attention.

5. Evaluation and Empirical Results

GTA-Link has set new state-of-the-art results in challenging sports tracking benchmarks. On SportsMOT, the HOTA metric increased for all trackers (e.g., Deep-EIoU: 77.21 → 81.04), with a marked reduction in ID switches. On SoccerNet and SoccerTrack, consistent HOTA improvements (e.g., ByteTrack: 67.30 → 71.97) were observed (Sun et al., 2024, Jian et al., 31 Jan 2026). Ablation studies show that the connector stage recovers most of the gain, while the splitter yields further incremental improvements, especially in scenarios with frequent mix-up errors.

A summary of observed improvements is presented below:

Tracker	HOTA Before	HOTA After	ID Switches Before	ID Switches After
SORT (SMOT)	56.28	66.52	5180	3547
ByteTrack	63.46	69.74	3147	2107
Deep-EIoU	77.21	81.04	2909	2737

GTA-Link results in negligible or modest increases in other error metrics such as false positives, while substantially boosting association accuracy. Similar patterns are observed for SoccerNet and SoccerTrack (Sun et al., 2024, Jian et al., 31 Jan 2026).

6. Algorithms in Other Domains and Methodological Variants

In multi-camera automotive tracking (Nguyen et al., 2022), GTA-Link employs a global association graph $M$9 in which nodes represent tracklets and new detections aggregated from multiple perspectives. Edges are scored by learned similarity and geometry features; link prediction employs a softmax over concatenated node features and pairwise descriptors, and attention mechanisms propagate both appearance and motion cues through the association process. Joint optimization of embedding and link-prediction losses ensures end-to-end learning of both node descriptors and association logic.

7. Limitations and Future Directions

GTA-Link is primarily limited by the quality of base appearance embeddings and the rigidity of hand-crafted spatial/temporal gating thresholds. Future avenues include the incorporation of motion-based costs, learnable spatial gating, and fully end-to-end, differentiable clustering objectives. In all published variants, GTA-Link remains a lightweight approach with runtime suitable for practical deployment, owing to the use of clustering and greedy merging rather than explicit integer programming (Sun et al., 2024, Jian et al., 31 Jan 2026).

---

For further details and open-source codebases, see (Sun et al., 2024, Nguyen et al., 2022), and (Jian et al., 31 Jan 2026).