Fréchet 始まり Distance (FID) is a metric used to evaluate the quality of images generated by 生成モデル, particularly in the context of 深層学習 and コンピュータビジョン. It is commonly employed to assess the performance of Generative Adversarial Networks (GANs) and other 画像合成 技術。
FIDは、2つの 確率分布: one representing the generated images and the other representing real images from a dataset. To compute FID, the images are first passed through a pre-trained Inception v3 neural network, which extracts feature representations of the images. These features are then modeled as multivariate Gaussian distributions, characterized by their mean and covariance.
FIDスコアは、これら2つの分布の距離を定量化するFréchet距離を用いて計算されます。これにより far apart these two distributions are. A lower FID score indicates that the generated images are more similar to the real images, suggesting better quality and diversity in the outputs of the generative model. Conversely, a higher FID score indicates poorer quality and greater divergence from real images.
全体として、FIDは堅牢な 評価指標です, allowing researchers and practitioners to compare various generative models effectively. It has become a standard benchmark in the field of generative modeling, helping to advance the development of high-quality image synthesis techniques.