Ilke Demir

Recent advances in the democratization of AI have been enabling the widespread use of generative models, causing the exponential rise of fake content. As every technology is simultaneously built with its counterpart to neutralize it, this is the perfect time to fortify our eyes with deep fake detectors. Deep fakes depend on photorealism to disable our natural detectors: we cannot simply look at a video to decide if it is real. On the other hand, this realism is not preserved in physiological, biological, and physical signals of deep fakes, yet. My key assertion follows that such signals hidden in portrait videos can be used as an implicit descriptor of authenticity, like a generalizable watermark of humans, because they are neither spatially nor temporally preserved in deep fakes.

One picture is worth a thousand words, so what have been told with videos? What about 100 simultaneous videos to reconstruct every frame of life in a 10.000 sq. ft dome? Similar to other industries, entertainment industry is also being reshaped by AI, especially towards AR/VR consumption. This talk will introduce deep learning advancements in 3D vision, reconstruction, and shape understanding with a focus on generative models. Then we will shift gears with an overview of such models in 3D, and their progression on voxels, point clouds, meshes, graphs, and other 3D representations. Back to our studio, in addition to a discussion about how to process such large visual data, the challenges of scaling 10x over current capture platforms, and over 200x over state-of-the-art datasets will be presented. The talk will conclude with a sneak peek of upcoming VR/AR productions from the world's largest volumetric capture stage at Intel Studios, as an example of real-world use cases.

Manual creation of massive and detailed 3D models may take weeks, or months, even for experienced artists. In contrast, generative algorithms enable efficiently creating content coherent with the reality, as long as the representations are good approximations of the real world. In this talk, I will introduce "proceduralization" to discover such generative representations from 2D and 3D data for synthesis, modeling, and reconstruction; combining computer vision, machine learning, and computational geometry approaches. I will briefly introduce geometry processing algorithms to discover artistic elements, and grammar discovery methods to extract procedural rules. The rest of the talk will be devoted to proceduralization applications for content creation, interactive editing, localization, mapping, and reconstruction. I will conclude by how similar constructions can be exploited in extracting shape abstractions in the context of geometric deep learning.