Coreset Selection for Object Detection via Multimodal LLMs

This project aims to develop a multimodal neural-network–based method for identifying a representative coreset from large-scale object detection datasets. The objective is to reduce dataset size while preserving the overall diversity, visual variability, and semantic richness necessary for training high-performance object detection models. The approach leverages a multimodal embedding space, where both images are encoded with a richer textual context.