Paper | ICML 2026 EMMQA Workshop | Dense-Set Dataset
LARE is a training-free augmentation for text-to-image retrieval in crowded scenes. Our method mines low-attention regions from a frozen vision encoder, encodes these regions alongside the full image, and combines regional embeddings with the global image embedding at inference time. This simple test-time procedure improves retrieval on Dense-Set variants that emphasize subtle and occluded content.
conda env create -f environment.yml
conda activate larefrom lare import create_lare
pipeline = create_lare(model="siglip-so400m")
score = pipeline.retrieve(image, "a person carrying a red bag")
print(score.score)Step 1. Clone & Environment
git clone https://github.com/aalquwayfili/LARE.git
cd LARE
conda env create -f environment.yml
conda activate lareStep 2. Download Datasets
Head to data/README.md for quick instructions and wget scripts on how to download the MS-COCO and Flickr30K image sets locally.
Step 3. Run Benchmark Scripts Pass the dataset directory to the reproduction script. It will evaluate the baseline and LARE pipeline natively against standard and Dense-Set annotations.
COCO_DIR=/path/to/coco/val2014 \
FLICKR_DIR=/path/to/flickr30k-images \
bash scripts/reproduce.sh siglip-so400m@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 Faisal Alhajari and Mohammed Alkhrashi
and Alreem Almuhrij and Abdullah Aldwyish and Raied Aljadaany
and Huda Alamri and Muhammad Kamran J. Khan},
booktitle={ICML 2026 Workshop on Efficient Multimodal Question Answering (EMM-QA)},
year={2026},
eprint={2606.18885},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://openreview.net/forum?id=42bo30qeLe}
}