MIRCaps: A Large-Scale Mixed-Domain Dataset with Image-Level and Region-Level Captions for Fine-Grained Vision-Language Learning

The article introduces MIRCaps, a large-scale multimodal dataset designed for fine-grained vision-language learning, featuring 141,364 images, 981,947 image-level captions, and 1,742,264 region-level captions with 1,391,779 bounding box annotations. This dataset aims to improve Vision-Language Models (VLMs) by providing diverse caption types that enhance the learning of visual attributes. Experimental results indicate that lightweight VLMs such as SmolVLM-256M-Instruct, BLIP, and Qwen2.5-VL 3B-Instruct can be effectively fine-tuned using MIRCaps, making it a valuable resource for practitioners in the field.