Perfect Perspective Correction with Artificial Intelligence

At Autoenhance.ai we’re working hard to create an artificial intelligence solution that levels human capabilities in enhancing real estate images. High quality enhancements are crucial to convey a property’s features and increase its perceived value. We believe artificial intelligence is able to fully automate these enhancements, which helps photographers to save costs and time. Everything we do is focused on improving the quality of real estate images. We’re constantly working on smarter AI models for every step of the pipeline, from merging HDR brackets to replacing gray skies with beautiful sunny ones. This time we dive deeper into the world of perspective correction. What is it and how are we automating it with AI?

The Significance of Perspective Correction in Real Estate Photography
‍
Perspective correction is a crucial aspect that significantly contributes to the visual aesthetics of the final image. If you take a real estate photo, it's often distorted because of the angle and position of the camera. Slightly tilted photos already largely affect how an image is perceived. To the human eye, images that have straight walls, doors and window frames for example, are way more pleasing to look at, because that’s how you naturally perceive a property as well. Therefore these images are being corrected afterwards, meaning that tilted or rotated walls are being straightened. One way human editors do this is by drawing lines over structural elements like walls in the image, which tells software how to straighten those images. This has to be done for every image that needs to be corrected, which costs a lot of time.

Navigating the Complexity of Perspective Correction
‍
Our goal to automate perspective correction with AI presents a multifaceted challenge. Autoenhance AI needs to understand the content of an image and identify elements that can be used for straightening. We also need to make sure that the AI is capable of finding these elements correctly every time and doesn’t get confused by elements that seem relevant but aren’t. To give an example, we know that walls, windows and doors are always built to be straight, so we can use those to correct the perspective. However, a tilted mirror, curtains in front of the windows or a curved chair can throw off an AI, resulting in incorrect perspectives.

Another factor that needs to be considered is the point of view the photographer has chosen. There are basically three points of view Autoenhance AI should learn to recognize and handle accordingly.

One point perspectives

One point perspectives, also known as "linear perspectives", are characterised by lines that converge at a single vanishing point on the horizon. In real estate photography, this is commonly seen when shooting straight-on facades of buildings or rooms. While it provides a straightforward and classic look, correcting perspective in one point images involves ensuring vertical lines remain truly vertical and horizontal lines are perfectly horizontal. This type of correction helps to prevent the image from appearing skewed or distorted.

Two point perspectives

Two point perspectives introduce a second vanishing point, creating a sense of depth and dimension, which occurs when rooms are photographed at an angle. It is also the most common point of view in real estate photography because it helps to show a room in the largest way possible. Correcting perspective in two point images mainly involves straightening the verticals (assuming the photographer hasn’t rotated his or her camera). Also straightening the horizontals would result in an inaccurate and warped correction, so Autoenhance AI needs to learn that it should disregard horizontal lines for these types of images.

Three point perspectives

In three point perspective images, vertical lines converge to a third vanishing point, typically positioned above or below the image frame. This is commonly observed when photographing from a tilted angle, emphasising height and scale. Correcting the perspective of these types of images would also result in a warped and distorted image.

How Autoenhance AI Learns to Find the Right Lines
‍
‍It’s becoming clear that automating perspective correction is a difficult task which involves a lot of training and visual understanding by our AI. To take on this challenge we have built a custom app that our human editors and experts use to draw lines over parts of the image that are relevant for correcting the perspective. Without giving away all our secrets, Autoenhance AI learns to detect these elements and decide for itself which ones it should use to correct the perspective. We have a very large dataset of interiors and exteriors, presenting enough opportunities for Autoenhance AI to learn which elements to take into account and which to disregard. Training AI models based on this data has already resulted in very promising corrected images.

Achieving Consistency Through Human Feedback
‍
One of our key goals is to provide consistency to our customers. We want you to know what you’re going to get when uploading an image to Autoenhance.ai. Therefore we’re constantly improving our perspective correction AI to get to 100% good results every time. One of the ways we’re going to do this is by enabling users to provide feedback to images that are not properly corrected. We then manually correct these images, feed this information back into the perspective correction AI and retrain it to get smarter. The combination of human input, an extensive dataset focused on real estate, and a constant feedback loop ensures a robust and adaptive AI system. We will continue improving Autoenhance AI models by looking at more and more data points, so in the end Autoenhance AI will be just as good as perceiving, correcting and enhancing images as a human would.