ChartREG++: Towards Benchmarking and Improving Chart Referring Expression Grounding under Diverse referring clues and Multi-Target Referring
Researchers have introduced ChartREG++, a new benchmark designed to improve the grounding of referring expressions in charts. This benchmark addresses limitations in existing datasets by supporting multiple localization forms, handling multi-instance references, incorporating diverse grounding cues, and covering a wider range of chart types. The team also developed a code-driven synthesis pipeline to generate pixel-accurate instance masks for training an instance segmentation model, which, when integrated into a multimodal grounding framework, demonstrated superior performance on their benchmark and generalized to real-world chart grounding tasks. AI
IMPACT Enhances multimodal models' ability to interpret and ground information within complex chart visualizations.