inkGround
A benchmark for stroke trajectory recovery from full-page, camera-captured handwritten images. It contains 537 page images, approximately 18,700 words and approximately 163,000 strokes, and is used to evaluate the fine-grained visual grounding ability of vision-language models on text-conditioned point localization tasks.
Timeline (last 90 days)
inkGround is a benchmark for stroke trajectory recovery on full-page, camera-captured handwritten images, containing 537 page images, about 18,700 words and about 163,000 strokes.
inkGround's reference trajectories are generated by a pretrained offline-to-online model and filtered for OCR consistency and verified by human raters, and are pseudo-labels.
inkGround defines the task as text-conditioned point localization: given a page image and a target word, the model must locate that word and output its ordered stroke points.
inkGround's evaluation metrics include trajectory fidelity (DTW and Chamfer distance), readability (Accurate Rate and Correct Rate) and style similarity.
The researchers benchmarked four open-source VLMs and three frontier closed-source VLMs using a shared contextual prompt.
Gemini-3.1-Pro stands out in trajectory fidelity and readability on the inkGround benchmark, with generated stroke trajectories approaching the reference level.
All Facts (6)
inkGround is a benchmark for stroke trajectory recovery on full-page, camera-captured handwritten images, containing 537 page images, about 18,700 words and about 163,000 strokes.
50%UnverifiedinkGround's reference trajectories are generated by a pretrained offline-to-online model and filtered for OCR consistency and verified by human raters, and are pseudo-labels.
50%UnverifiedinkGround defines the task as text-conditioned point localization: given a page image and a target word, the model must locate that word and output its ordered stroke points.
50%UnverifiedinkGround's evaluation metrics include trajectory fidelity (DTW and Chamfer distance), readability (Accurate Rate and Correct Rate) and style similarity.
50%UnverifiedGemini-3.1-Pro stands out in trajectory fidelity and readability on the inkGround benchmark, with generated stroke trajectories approaching the reference level.
50%UnverifiedThe researchers benchmarked four open-source VLMs and three frontier closed-source VLMs using a shared contextual prompt.
50%