File size: 2,874 Bytes
2d66b28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
---
license: mit
pipeline_tag: video-text-to-text
tags:
- multi-object-tracking
- video-understanding
- vision-language-model
- spatiotemporal-reasoning
---

# QTrack: Query-Driven Reasoning for Multi-modal MOT

[**QTrack**](https://huggingface.co/papers/2603.13759) is an end-to-end vision-language model designed for query-driven multi-object tracking (MOT). Unlike traditional MOT which tracks all objects in a scene, QTrack selectively localizes and tracks specific targets based on natural language instructions while maintaining temporal coherence and identity consistency.

- **Paper:** [QTrack: Query-Driven Reasoning for Multi-modal MOT](https://huggingface.co/papers/2603.13759)
- **Project Page:** [https://gaash-lab.github.io/QTrack/](https://gaash-lab.github.io/QTrack/)
- **Repository:** [https://github.com/gaash-lab/QTrack](https://github.com/gaash-lab/QTrack)

## Description

Multi-object tracking has traditionally focused on estimating trajectories of all objects. QTrack introduces a **query-driven tracking paradigm** that formulates tracking as a spatiotemporal reasoning problem conditioned on natural language queries. 

### Key Contributions
- **RMOT26 Benchmark**: A large-scale benchmark with grounded queries and sequence-level splits to enable robust evaluation of generalization.
- **QTrack Model**: An end-to-end vision-language model that integrates multimodal reasoning with tracking-oriented localization.
- **Temporal Perception-Aware Policy Optimization (TPA-PO)**: A structured reward strategy to encourage motion-aware reasoning.

## Benchmark Results

QTrack achieves state-of-the-art performance on the [RMOT26](https://huggingface.co/datasets/GAASH-Lab/RMOT26) benchmark.

| Model | Params | MCP↑ | MOTP↑ | CLE (px)↓ | NDE↓ |
|:-----:|:------:|:----:|:-----:|:---------:|:----:|
| GPT-5.2 | - | 0.25 | 0.61 | 94.2 | 0.55 |
| Qwen3-VL-Instruct | 8B | 0.25 | 0.64 | 96.0 | 0.97 |
| Gemma 3 | 27B | 0.24 | 0.56 | 58.4 | 0.88 |
| InternVL | 8B | 0.21 | 0.66 | 117.44 | 0.64 |
| **QTrack (Ours)** | **3B** | **0.30** | **0.75** | **44.61** | **0.39** |

## Installation

To set up the environment and use the model, please follow the instructions in the [official repository](https://github.com/gaash-lab/QTrack):

```bash
# Create conda environment
conda create -n qtrack python=3.12
conda activate qtrack

# Install QTrack and dependencies
git clone https://github.com/gaash-lab/QTrack.git
cd QTrack
pip install -r requirements.txt
pip install -e .
```

## Citation

If you find QTrack useful for your research, please cite:

```bibtex
@article{ashraf2026qtrack,
  title={QTrack: Query-Driven Reasoning for Multi-modal MOT},
  author={Ashraf, Tajamul and Tariq, Tavaheed and Yadav, Sonia and Ul Riyaz, Abrar and Tak, Wasif and Abdar, Moloud and Bashir, Janibul},
  journal={arXiv preprint arXiv:2603.13759},
  year={2026}
}
```