AI PhotoRobot

About the project

The 3 main and perhaps most popular areas of AI are image and text/speech recognition, as well as information evaluation. A key element of these systems is the creation of the algorithms that perform these tasks. This step may be uninteresting to end users, but it's very important for AI developers – these algorithms are no longer created manually by developers, but by systems from Google, Microsoft, or IBM (TensorFlow, Azure, Watson, etc.), which require an enormous amount of base data to be provided. For this, texts and data are usually available, such as medical records and confirmed diagnoses, or even a city's public transit data. In the case of machine vision and image recognition, this means thousands of photos in which the element to be recognized – image, object, logo, etc. – is definitely present. These images must all be completely different from each other. If, for example, a manufacturing company wants to develop an AI solution to recognize any of its products this way, creating even hundreds of thousands of such photos is horrendously tedious, manual work. Our PhotoRobot image generator offers a solution to this problem. The “image generator” works very simply. From even a single photo of the object to be recognized, we can generate the thousands of images needed for training – each different in every respect – within minutes. All that's needed is a 2D photo, or photos, of the object to be recognized. If we want to define the environments in which the object should appear, we can provide these ourselves; if this isn't important for training purposes, we use our own image database instead. The Image Generator is a piece of software that, depending on the parameters set, can place the key object in front of a random background image in a fraction of a second, in a random position and orientation, under different lighting, blurred or partially obscured – then creates and saves an appropriately sized image of this. It repeats this process thousands of times in a row, dramatically speeding up the time required to produce the images needed for training. Currently, we've built this software for our own use only, but we plan to make it available soon as a web-based service as well.