---
license: mit
library_name: DETECTRON2
tags:
- computer-vision
- image-segmentation
- video-instance-segmentation
- deep-learning
- query-enrichment
- pytorch
datasets:
- youtube-vis
- ovis
pipeline_tag: image-segmentation
metrics:
- ap
- ar
base_model:
- deahuang/MinVIS-Model-Zoo
---

# QueenVIS: Rethinking Image-Only Training for Video Instance Segmentation via Query Enrichment

This repository hosts the official model weights and evaluation benchmarks for **QueenVIS**.

## 🔗 Quick Links

* **Project Page:** [aut-aisl.github.io/QueenVIS-Project-Page](https://aut-aisl.github.io/QueenVIS-Project-Page/)
* **GitHub Repository:** [github.com/ArianKheir/QueenVIS](https://github.com/ArianKheir/QueenVIS)
* **Paper (arXiv):** [10.48550/arXiv.2607.24598](https://doi.org/10.48550/arXiv.2607.24598)

---

## 📖 Model Description

**QueenVIS** rethinks the paradigm of image-only training for Video Instance Segmentation (VIS). By introducing training-only auxiliary heads for feature prediction and center prediction, QueenVIS embeds dense appearance and spatial priors directly into transformer object queries without requiring video clip supervision.

* **Key Feature:** Adds **zero parameters** and **zero extra GFLOPs** at inference relative to image-only baselines like MinVIS.
* **Temporal Association:** Employs a training-free confidence-guided query propagation mechanism paired with a non-parametric memory bank to maintain temporal identity across frames.

---

## 📊 Datasets & Evaluation Metrics

QueenVIS is evaluated across standard VIS benchmarks:

* **YouTube-VIS (2019 / 2021 / 2022 Long Split):** Standard and long-sequence video benchmarks.
* **OVIS (Occluded Video Instance Segmentation):** A challenging benchmark characterized by severe occlusion and complex object interactions.

Performance is measured using standard Video Instance Segmentation metrics:
* **AP (Average Precision):** Primary metric across IoU thresholds.
* **AP₅₀ & AP₇₅:** Average Precision at IoU thresholds of 0.50 and 0.75.
* **AR₁ & AR₁₀:** Average Recall given 1 and 10 detections per video.

---

## 📈 Evaluation Results

### 1. Main Benchmarks (ResNet-50 Backbone)

| Dataset | AP | AP₅₀ | AP₇₅ | AR₁ | AR₁₀ |
| :--- | :---: | :---: | :---: | :---: | :---: |
| **YouTube-VIS 2019** | **51.8** | 75.2 | 58.3 | 49.4 | 62.1 |
| **YouTube-VIS 2021** | **50.9** | 72.1 | 55.8 | 42.8 | 57.3 |
| **YouTube-VIS 2022 (Long)** | **33.6** | 56.6 | 34.0 | 31.7 | 39.9 |
| **OVIS (Occluded VIS)** | **29.8** | 52.7 | 29.0 | 15.4 | 34.6 |

---

### 2. Scaling to Vision Transformers (Swin-L Backbone)

| Dataset | AP | AP₅₀ | AP₇₅ | AR₁ | AR₁₀ |
| :--- | :---: | :---: | :---: | :---: | :---: |
| **YouTube-VIS 2019** | **63.2** | 85.3 | 68.3 | 55.5 | 68.4 |
| **YouTube-VIS 2021** | **59.8** | 81.5 | 65.2 | 48.3 | 64.2 |
| **OVIS (Occluded VIS)** | **41.0** | 62.2 | 43.0 | 18.4 | 45.0 |

---

## 🚀 How to Use

For complete instructions on installation, environment setup, dataset preparation, and inference scripts, visit the official [GitHub Repository](https://github.com/ArianKheir/QueenVIS).

---

## 📚 Citation

If you find QueenVIS useful in your research, please cite our work:

```bibtex
@misc{kheirandish2026queenvis,
      title={QueenVIS: Rethinking Image-Only Training for Video Instance Segmentation via Query Enrichment}, 
      author={Arian Kheirandish and Fardin Ayar and Ehsan Javanmardi and Manabu Tsukada and Mahdi Javanmardi},
      year={2026},
      eprint={2607.24598},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={[https://doi.org/10.48550/arXiv.2607.24598](https://doi.org/10.48550/arXiv.2607.24598)}
}