ChartREG++: Towards Benchmarking and Improving Chart Referring Expression Grounding under Diverse referring clues and Multi-Target Referring

The article presents ChartREG++, a new benchmark for chart referring expression grounding that addresses limitations in existing benchmarks by supporting multiple localization forms, handling multiple referred targets, and encompassing a wider variety of chart types. The authors introduce a code-driven synthesis pipeline for generating pixel-accurate instance masks, which enhances an instance segmentation model integrated into a multimodal grounding framework. This advancement reveals significant performance gaps in current models and demonstrates improved generalization to real-chart grounding tasks, providing practitioners with a more robust tool for visual grounding in complex chart scenarios.