Intel Labs shows off new enhanced photorealism techniques

Video game developers are always looking for more realism in their games. The more realistic images look, the more engaging and immersive video games can be, particularly in virtual reality. Researchers from Intel Labs have presented a new approach to enhancing the realism of synthetic images. The images are enhanced by a convolutional network designed to leverage intermediate representations produced by conventional rendering pipelines.

Researchers trained the network using an adversarial objective that provides strong supervision at multiple perceptual levels. By analyzing scene layout distributions and commonly used data sets, the researchers found they differed in significant ways. The researchers on the project hypothesized that these differences are one of the causes of artifacts observed in many prior methods for enhancing photorealism.

To get around the issue of artifacts, the scientists have proposed a new strategy for sampling image patches while training the network. The team also leveraged multiple architectural improvements in deep network modules that are leveraged for photorealism enhancement. The method the researchers have developed certainly shows significant improvements. In the image above, the videogame GTA V is shown on the left, and the resulting photorealistic enhancement from the researchers at Intel Labs is shown on the right.

The scene on the right side of the composite image looks like it was taken with a dash camera, while the left side of the screen shows an attractive image that is clearly a videogame. The same enhancements also make landscapes and cityscapes much more realistic. Trees, grass, and dirt paths look just as they would in the real world rather than the somewhat cartoonish appearance they take on in GTA V.

The system developed by Intel Labs does a fantastic job of taking scenes from GTA V and making them much more realistic. It's impressive how much of an improvement the roadways gain.