Researchers create robotic 'photographer' that knows how to choose a good composition to take pictures

Researchers from Cornell University have designed a photography robot, which can understand harmonious aesthetic composition. It is initially trained for interior photography, but can be trained further to use the skill anywhere.

This photography robot is called AutoPhoto.

The robot is called AutoPhoto and was developed by graduate student Hadi AlZayer, along with two researchers Hubert Lin and Kavita Bala. AlZayer said the idea came from the desire to have a companion who could take photos from a better angle than selfies or when asking someone else to take them.

'Whenever I ask strangers to take pictures of me, I get poorly composed photos,' he says. This prompted AlZayer to find a way to create a photography robot.

What people think to create a photo with aesthetic composition is complex, but AlZayer believes it can be automated with algorithms. Using a pre-generated basic algorithm, he helps refine the technique through a process of 'reinforcement learning', which is a subfield of machine learning.

This is the first robotic system to be "taught" an aesthetic machine learning model.

AutoPhoto is said to be the first robotic system to be "taught" with an available aesthetic machine learning model, and represents a major development in the use of automated robots for photography. This robot can know what makes a good photo.

The aesthetic machine learning model is trained on a dataset of more than a million photos that have been aesthetically graded by humans. AutoPhoto itself also trained on dozens of 3D images of interior scenes in the room to determine the optimal composition angle, before the team attached it to Jackal, a small 4-wheeled robot, to move.

The robot knows how to choose a good composition when taking photos.

When dropping it on a campus building, as you can see in the video above, the robot starts taking a series of bad composition photos, but as the AutoPhoto algorithm gets used to the environment, the composition selection Its image was steadily improved until the image matched the advertising photos of real estate companies. The robot takes about a dozen iterations on average to optimize each scene, but the whole process only takes a few minutes to complete.

On average, the robot takes about a dozen repeat shots to optimize each scene.

In the future, AlZayer hopes to adapt the AutoPhoto system for outdoor use, potentially swapping the Jackal and drone. 'Capturing high-quality outdoor scenes is difficult,' says AlZayer, 'because it's harder to control composition'.

The researchers believe this is just the beginning of automated photography, allowing robots to capture beautiful images of dangerous or remote environments without human intervention.