Advancing Image-to-Image Translation Model Assessment with Clustering Technique
Rung-Ching Chen, Chayanon Sub-r-pa · 2024
The Fréchet Inception Distance (FID) is a metric commonly used to assess the quality of images generated by generative models. However, FID may not be suitable for evaluating Image-to-Image (I2I) translations since the results may sometimes have defects or may not be in the target domain. We proposed a new method to assess I2I translations, which involves calculating a new scoring system. Single Image Inception Distance (SiID), a modified version of FID, assigns a quality score to each generated image and CNN-based classification model to determine its classification probability. SiID and classification probability are then used in a clustering algorithm to gain more insight into evaluating the I2I Model. Our proposed method has been demonstrated on existing I2I translation models for translating male-to-female images, and the results have shown that our method can classify the generated images into three groups: successfully translated, lacking realism, and failing to translate.