Inception Score (IS)

Inception Score (IS) is an objective performance metric used to evaluate the quality of generated images or synthetic images generated by a generative adversarial network (GAN). It can measure the authenticity and diversity of the output images and can also be used to replace human subjective evaluation. IS is the second most important evaluation performance metric after FID (Frechlet Inception Distance).

This concept was first proposed by Tim Salimans et al. in their 2016 paper "Improved Techniques for Training GANs"It was introduced in . It is named after the Inception Network (a deep learning model pre-trained on the Image net dataset) developed by Google for image classification. The author used the Inception Network to extract the feature vector of the image.

It measures two aspects:

• Variety — There should be a lot of variety in the generated images — the entropy of the overall distribution should be high.
• Goodness — how good the generated image is — requires low entropy and high predictability.

The lowest IS can be zero, the highest IS can be infinity, and the higher the IS, the better.

Limitations of Inception Score

1. If the network only generates one image per class, the IS will be high and will not represent the class well.
2. The Inception network was trained on the Imagenet dataset which only contains 1,000 classes. If you train a GAN on classes outside of these 1,000, your Inception score will be low.
3. It works well with images that are square and relatively small in size – for example, 300×300 in size.

References

【1】https://kailashahirwar.medium.com/a-very-short-introduction-to-inception-score-is-c9b03a7dd788