Post Quality

Outlier Detection

BaseLine - CLIP Score Threshold

A natural idea is to threshold based detection over CLIP score.

Two Sided thresholding

Clipping

Rescaling

$$r(\mathbf{x_i}) = \left(1 - \left[\frac{1}{1 + e^{-k_{LB}(x - x_{LB})}} - \frac{1}{1 + e^{-k_{UB}(x - x_{UB})}} \right] \right)s(\mathbf{x}_i,f_\theta)$$

CLIP Score (CIFAR 100)

(Ours) PQ Score (CIFAR 100)

CLIP Score (CIFAR-10)

CIFAR 100

Metric	Precision	Recall	F1 Score	Accuracy
CLIP Score	99.94	95.71	97.78	95.70
PQ Score	99.93	96.71	98.30	96.67

CLIP Score

(Ours) PQ Score

References:

• CLIPScore: A Reference-free Evaluation Metric for Image Captioning

• ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation