Fine-grained template sub-category classification

Develop methods that substantially improve sub-category classification of rendered graphic design templates within the LICA/GDB taxonomy, which currently achieves only 10.13% Top-1 accuracy even under label-constrained prompting.

Background

The paper evaluates template category prediction at two levels—parent categories and sub-categories—under open-vocabulary and label-constrained prompting. While constraining labels improves parent-category accuracy, sub-category classification remains very low.

The authors explicitly characterize sub-category classification as an open challenge, emphasizing that performance is poor even when the label set is provided, indicating a gap that current models do not bridge.