Precise Numerical Recovery from Chart Images

Develop a dedicated and robust mechanism for precise numerical recovery from chart images, including reliable extraction of data values and the alignment of recovered values with chart geometry, across diverse and customized chart layouts.

Background

Chart2SVG evaluates numerical accuracy by parsing the geometry of reconstructed SVGs with DIVI-based heuristics rather than relying on a vision-language-model extraction oracle. The paper explains that current VLMs achieve only moderate accuracy when reading values directly from chart images, while existing chart-to-table systems are often limited to charts with explicit annotations or standard layouts.

The authors state that current chart-to-SVG and chart-to-code systems lack a dedicated robust mechanism for precise numerical recovery. They identify tighter integration of OCR-based recovery—either as a post-hoc module or as an architectural component—as a future direction, because reliable axis-scale inference and value-to-geometry mapping remain unresolved across diverse charts.

References

We adopt this rule-based approach because numerical value extraction remains an open challenge across the field: even state-of-the-art VLMs such as o3 and Gemini-2.5-Flash achieve only moderate accuracy when reading data values directly from chart images, and dedicated chart-to-table models are typically restricted to charts with explicit, unambiguous value annotations or standard layouts, limiting their applicability to the diverse and customized charts in our evaluation.

— Chart2SVG: Editable SVG Generation from Raster Chart Images  (2608.26544 - Cui et al., 27 Aug 2026) in Appendix, Section “Baseline Settings and Comparative Analysis,” subsection “Data Value Extraction Methodology”

However, this would likely require recovering the underlying data and legend specification from rendered charts, which is a more challenging problem. Moreover, the extent to which such recovered representations can support precise and semantically consistent interventions remains unclear.

— LegendBench: A Diagnostic Benchmark for Legend Understanding with Counterfactual Interventions  (2609.24172 - Zhang et al., 21 Sep 2026) in Discussion, paragraph “Diversity and Control”