Dense-Set: Visually Dense Scene Benchmark for Text-to-Image Retrieval

Dense-Set is a curated benchmark for evaluating text-to-image retrieval systems on challenging, visually dense scenes. It was created by AbdulmalekDS and published in 2026 alongside research presented at the ICML 2026 Workshop on Efficient Multimodal Question Answering. The benchmark comprises subsets extracted from the COCO and Flickr30K datasets.

Use Cases

• Benchmarking text-to-image retrieval models based on performance on crowded images with multiple object instances.
• Evaluating model robustness based on retrieval accuracy for underrepresented, low-attention object classes.
• Training or fine-tuning multimodal models based on challenging subsets from established datasets like COCO and Flickr30K.

Strengths

• Focuses on a specific, challenging task: text-to-image retrieval on visually dense scenes.
• Curated from established, widely-used source datasets (COCO and Flickr30K).
• Published alongside peer-reviewed research (ICML 2026 Workshop).

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.

Provenance

Source

Subsets extracted from COCO and Flickr30K datasets.

Collection Method

Curated selection focusing on crowded images and low-attention classes.

Freshness

Last updated 2026-06-17 06:23:17; freshness should be verified.