Add all 5,566 images, clean up columns, integrate CVPR 2026 MULA statistics

README.md

@@ -14,21 +14,34 @@ dataset_info:
 
 **Dense-Set** is a benchmark dataset tailored for text-to-image retrieval in densely crowded scenes. It provides hard, fine-grained, and low-attention subsets extracted from the MS-COCO and Flickr30k validation splits, intentionally emphasizing rare classes that traditional vision-language models frequently overlook.
 
-This dataset is the official benchmark published alongside our paper:
+This dataset is the official benchmark published alongside our MULA CVPR 2026 paper:
 
 > **LARE: Low-Attention Region Encoding for Text–Image Retrieval**<br>
-> *Accepted at the [MULA Workshop](https://mula-workshop.github.io/)*
+> *Accepted at the [CVPR 2026 MULA Workshop](https://mula-workshop.github.io/)*
 
-## Dataset Subsets
+## Dataset Statistics
 
-You can explore the two individual subsets cleanly separated using the dropdown menu at the top of the Hugging Face Dataset Viewer!
+The LARE Dense-Set specifically isolates crowded scenes with complex local attributes. Our extraction statistics from the parent datasets are as follows:
+
+| Dataset | Split | # Images | Avg. Objects | Avg. # Classes |
+| :--- | :--- | :--- | :--- | :--- |
+| **MS-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** |
+
+## Subsets & Pre-Loaded Images
+
+You can explore the two individual subsets cleanly separated using the dropdown menu at the top of the Hugging Face Dataset Viewer! All subset images are natively hosted on this repository, allowing visual exploration of the `rare_classes` mapping algorithms directly on the Hub.
 
 - **`coco`**: 3,089 images with dense captions and localized rare-class mappings.
 - **`flickr30k`**: 2,477 images with dense captions and localized rare-class mappings.
 
 ### Loading the Data
 
-Hugging Face's `datasets` library handles the config subsets natively:
+Hugging Face's `datasets` library handles downloading the targeted metadata configs natively:
 
 ```python
 from datasets import load_dataset
@@ -41,9 +54,3 @@ flickr_ds = load_dataset("AbdulmalekDS/Dense-Set", "flickr30k")
 
 print(coco_ds["test"][0])
 ```
-
-## Images
-
-In accordance with standard multimodal dataset practices, **we do not host the raw gigabyte image archives**. The images referenced in this dataset (`filename` column) correspond to the standard **MS-COCO val2014** and **Flickr30k** image sets. 
-
-Researchers evaluating on Dense-Set should download the standard image corpus from their original maintainers and utilize our JSON Lines metadata to evaluate retrieval configurations.