By Enric Moreu, ESR at Dublin City University
Artificial intelligence requires large amounts of data to train a neural network. Specifically, some computer vision algorithms need millions of images to reach a good performance. In general, those images are only useful when they are labelled. In other words, a human has to work on the expensive and slow task of annotating images.
In that context, my current research consists in exploring how synthetic images can help the computer vision algorithms.
Synthetic data is any content artificially generated that mimics the data gathered from the real world. Synthetic data does not require a human labeling images because the system knows precisely the properties of the generated content.
Currently, I am exploring how synthetic data can improve the performance of the object counting task. Object counting consists in identifying the number of cells in a medical image, the number of people in a stadium or the number of cars on a road. Using the latest 3D technology I create synthetic datasets that help AI to have a better understanding of reality.
While humans spend around 30 seconds to count the number of objects in a real-world image, machines can generate a synthetic image with an extremely detailed knowledge of the items in the image. In terms of costs, a human annotator earns around €0.2 per image whereas the computing cost of generating a synthetic one is nearly zero.
The core of my research relies on a tool that I developed to rapidly generate synthetic images based on several parameters such as the background, the lighting and the 3D models . It is highly scalable and allows the production of hundreds of realistic images every second.