Incorporate methodological summary strictly matching the paper draft

README.md

@@ -1,66 +1,20 @@
----
-configs:
-- config_name: coco
-  data_files:
-  - split: test
-    path: coco/test-*
-- config_name: flickr30k
-  data_files:
-  - split: test
-    path: flickr30k/test-*
-dataset_info:
-- config_name: coco
-  features:
-  - name: image
-    dtype: image
-  - name: split
-    dtype: string
-  - name: sentences
-    list: string
-  - name: sentids
-    list: int64
-  - name: rare_classes
-    list:
-      list: string
-  splits:
-  - name: test
-    num_bytes: 588606142
-    num_examples: 3089
-  download_size: 575320921
-  dataset_size: 588606142
-- config_name: flickr30k
-  features:
-  - name: image
-    dtype: image
-  - name: split
-    dtype: string
-  - name: sentences
-    list: string
-  - name: sentids
-    list: int64
-  - name: rare_classes
-    list:
-      list: string
-  splits:
-  - name: test
-    num_bytes: 362160915
-    num_examples: 2477
-  download_size: 375495886
-  dataset_size: 362160915
----
-
 # LARE: Dense-Set
 
-**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.
+**Dense-Set** is a curated benchmark of visually dense scenes tailored for text-to-image retrieval. It provides challenging evaluation subsets extracted from MS-COCO and Flickr30K, containing crowded images with multiple object instances and underrepresented, low-attention classes.
 
-This dataset is the official benchmark published alongside our MULA CVPR 2026 paper:
+This dataset is the official benchmark published alongside our paper:
 
 > **LARE: Low-Attention Region Encoding for Text–Image Retrieval**<br>
 > *Accepted at the [CVPR 2026 MULA Workshop](https://mula-workshop.github.io/)*
 
-## Dataset Statistics
+## Dataset Construction & Re-captioning
+
+Dense-Set was constructed through an automated pipeline carefully designed to emphasize objects traditional vision-language models overlook:
+1. **High-Density Filtering**: Images were processed using a YOLO object detector. We ranked images by total object count and isolated the top 10% to create a high-density candidate pool.
+2. **Rare-Class Isolation**: Within the dense pool, we identified "rare classes" at the image level—defined as object categories appearing exactly once in a given image. 
+3. **BLIP-2 Re-captioning**: To shift the textual focus away from general scene context, we filtered out rare-class detections occupying >15% of the image area. We then prompted BLIP-2 with class-aware templates to explicitly describe these small or underrepresented objects, producing highly challenging, fine-grained captions.
 
-The LARE Dense-Set specifically isolates crowded scenes with complex local attributes. Our extraction statistics from the parent datasets are as follows:
+## Dataset Statistics
 
 | Dataset | Split | # Images | Avg. Objects | Avg. # Classes |
 | :--- | :--- | :--- | :--- | :--- |
@@ -71,16 +25,9 @@ The LARE Dense-Set specifically isolates crowded scenes with complex local attri
 | | *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
 
-### Loading the Data
-
-Hugging Face's `datasets` library handles downloading the targeted metadata configs natively:
+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 within a Parquet structure.
 
 ```python
 from datasets import load_dataset
@@ -93,3 +40,16 @@ flickr_ds = load_dataset("AbdulmalekDS/Dense-Set", "flickr30k")
 
 print(coco_ds["test"][0])
 ```
+
+## Citation
+
+*Our camera-ready paper will be available shortly. Please use the following citation placeholder in the meantime:*
+
+```bibtex
+@inproceedings{lare2026,
+  title={LARE: Low-Attention Region Encoding for Text–Image Retrieval},
+  author={Anonymous},
+  booktitle={CVPR MULA Workshop},
+  year={2026}
+}
+```