1. 성능 평가 개요

• Data → Train/Test split → Validation Split → K-Fold Cross Validation

• Confusion Matrix

• Recall Precision

2. OCR model evaluation methods

2-1. DetEval

Notation

• A list $G$ of ground truth object rectangles $G_i \ (i = 1 \dots |G|)$
• A list $D$ of detected object rectangles $D_j \ (j=1 \dots |D|)$
• Area Recall $R_{AR}$

$$ \sigma_{ij} = R_{AR}(G_i,D_j) = \frac{Area(G_i \cap D_j)}{Area(G_i)} $$

• Area Precision $P_{AR}$

$$ \tau_{i,j} = P_{AR}(G_i,D_j) = \frac{Area(G_i \cap D_j)}{Area(D_i)} $$

• Overlap matrices $\sigma$ and $\tau$
• $t_r \in [0,1]$ is the constraint on area recall
• $t_p \in [0,1]$ is the constraint on area precision