README.md

---
license: cc-by-4.0
configs:
  - config_name: coco
    data_files:
      - split: test
        path: coco/test-*
  - config_name: flickr30k
    data_files:
      - split: test
        path: flickr30k/test-*
---

# Dense-Set

**Dense-Set** is a curated benchmark of visually dense scenes for text-to-image retrieval evaluation. It provides challenging subsets extracted from COCO and Flickr30K, focusing on crowded images with multiple object instances and underrepresented, low-attention classes.

This dataset is published alongside:

> **LARE: Low-Attention Region Encoding for Text–Image Retrieval**
> *CVPR 2026 — [MULA Workshop](https://mula-workshop.github.io/)*
> [Project Page](https://falmeshal.github.io/LARE/) | [Code](https://github.com/AbdulmalikDS/LARE)

## Dataset Samples

![Dense-Set samples](assets/samples.png)
*Each image is re-captioned to explicitly describe rare or low-attention objects (highlighted in red), shifting focus away from dominant scene context.*

## Construction

Dense-Set was built through a three-stage pipeline designed to surface objects that standard vision-language models overlook:

1. **High-Density Filtering** — Images processed with YOLO, ranked by total object count, top 10% retained as the high-density candidate pool.
2. **Rare-Class Isolation** — Within the dense pool, object categories appearing exactly once per image are flagged as rare classes, corresponding to small or visually subordinate objects.
3. **Re-captioning** — Rare-class detections occupying >15% of the image are filtered out. BLIP-2 is prompted with class-aware templates to explicitly describe the remaining underrepresented objects, producing fine-grained captions that shift focus away from dominant scene context.

## Statistics

| Dataset | Split | # Images | Avg. Objects | Avg. # Classes |
| :--- | :--- | ---: | ---: | ---: |
| **COCO** | Original Test Set | 40,504 | 6.71 | 2.85 |
| | High-Density Subset | 4,050 | 21.63 | 4.82 |
| | **Dense-Set** | **3,089** | **21.63** | **5.47** |
| **Flickr30K** | Original Test Set | 31,783 | 6.73 | 2.48 |
| | High-Density Subset | 3,178 | 19.40 | 4.38 |
| | **Dense-Set** | **2,477** | **19.55** | **4.85** |

## Usage

```python
from datasets import load_dataset

coco_ds   = load_dataset("AbdulmalekDS/Dense-Set", "coco")
flickr_ds = load_dataset("AbdulmalekDS/Dense-Set", "flickr30k")

print(coco_ds["test"][0])
```

## Acknowledgements

Dense-Set is built on images from [COCO](https://cocodataset.org) and [Flickr30K](https://shannon.cs.illinois.edu/DenotationGraph/).

## Citation

```bibtex
@inproceedings{alquwayfili2026lare,
  title={LARE: Low-Attention Region Encoding for Text--Image Retrieval},
  author={Abdulmalik Alquwayfili and Faisal Almeshal and Jumanah Almajnouni and Leena Alotaibi and Huda Alamri and Muhammad Kamran J Khan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
  year={2026}
}
```