Tournament-style Image Rating Tool

Selecting the perfect teaser image from a huge collection of images is a difficult and very time consuming task. There are many different factors that are considered by experts when selecting images for further use: From the image composition, the mood, the camera position, colors, objects in the scene and facial expressions to whether the frame is representative of the story.

In this project we devise methods for the automatic discovery of aesthetically pleasing images. For this reason, we design a labeling tool for our industry partners to rate images according to their own tastes, so that we can reproduce their picks and save them time. The editors are asked to indicate which of two simultaneously presented images they prefer aesthetically (“Which image do you like better?”). The compared images always belong to the same query so that they are similar in content and the labeller can focus on the aesthetics only.

The collected data allows us not only to obtain scores per image to train our models, but also to analyze correlations between labellers and detect outliers.

The code, together with user instructions can be found here: external page https://github.com/MTC-ETH/labelling-tool-aesthetics