Facebook's Artificial Intelligence Experts Claims They Have Developed A Better Way To Create 3D Images

According to a group of artificial intelligence (AI) researchers from the National Tsing Hua University in Taiwan, Facebook, and Virginia Tech, they have developed a better way to create 3D images. The researchers state that their method is superior to Facebook's current 3D Photos feature and other already existing methods to generate 3D photos.

Facebook rolled out its 3D Photos back in 2018 for dual-camera mobiles such as iPhone X. Facebook 3D Photos uses TrueDepth camera to evaluate the depth in images. In the latest study, the researchers use several images captured with an iPhone to indicate how their approach eliminates the discontinuity and blur other existing 3D methods offer.

This new way may also make for better Facebook 3D Images, however, if the new method translates to other environments, it may also result in more real immersion in environments with 3D graphics.

This new way is able to create 3D images from RGB-D imagery and it also operates with regular 3D images by using a pre-trained depth evaluation system. The research work also claims superior performance than LLFF and Nvidia’s Xview. Researchers assessed the performance of some 3D models by using imagery from the RealEstate 10K data set. During the past few months, Nvidia, Microsoft, and Facebook have introduced technologies to create 3D objects from 2D photos. However, the new method uses inpainting, a process of artificial intelligence predicting missing pixels in photos, to generate 3D images.

The new cutting-edge 3D image approach was published in a paper on preprint arXiv. According to the authors, their approach is different as it relies heavily on inpainting for depth value as well as color predictions. Moreover, the new method does not also require pre-determining a specific number of layers for generating 3D photos.


The authors have validated their approach to a wide variety of photographs clicked in various distinct situations. The IEEE Conference on CVPR (Computer Vision and Pattern Recognition) accepted the research paper for publication. It will be published next month. The Artificial Intelligence Index 2019 report shows that CVPR is one of the most significant annual machine learning conferences for artificial intelligence research experts.



Read next: Facebook Introduces AI to Make eCommerce Easier

No comments:

Post a Comment