Equally critical, with our technology, 3D artists in retail, media and entertainment, and other industries pressured to deliver more—and more immersive AR—experiences can reduce costs and speed to generate photorealistic 3D models, as much as tenfold. We know this from our own work because we’ve seen computing costs to generate the highest-quality 3D experiences drop significantly—even though we run an advanced Compute Engine loaded with a powerful GPUs, high-end CPUs, and massive amounts of RAM. If the goal is to scale industry-leading compute power quickly for a global customer base, Google Cloud is the proper solution.
Cloud Storage is another key but often overlooked component of the Google Cloud ecosystem, critical for 3co. We need the high throughput, low latency, and instant scalability delivered bylocal cloud SSDs to support the massive amounts of data we generate, store, and stream. The local SSDs complement our A2 compute engines and are physically attached to the servers hosting the virtual machine instances. This local configuration supports extremely high input/output operations per second (IOPS) with very low latency compared to persistent disks.
To top it off,Cloud Logging delivers us real-time log management at exabyte scale — ingesting analytic events that are streamed to data lakes withPub/Sub – so we can know while enjoying the beach here in Miami, Florida that everything is going smoothly in the cloud.
Building the 3co AI stack with TensorFlow
Building one of the world's most advanced 3D computer vision solutions would not have been possible withoutTensorFlow and its comprehensive ecosystem of tools, libraries, and community resources. Since the launch of TensorFlow in 2015, I’ve personally built dozens of deep learning systems using this battle-hardened technology, an open source Google API for AI. Through TensorFlow on Google Cloud, 3co is able to scale its compute power for creation of truly photorealistic digital models of physical objects — down to microscopic computation of material textures, and deep representations of surface light transport from all angles.
Most recently, 3co has been making massive progress on top of the TensorFlow implementation of Neural Radiance Fields (“NeRF”, Mildenhall et al. 2020). We are humbled to note that this breakthrough AI in TensorFlow truly is disruptive for the 3D modeling industry: we anticipate the next decade in 3D modeling will be increasingly shaped and colored by similar neural networks (I believe the key insight of the original authors of NeRF is to force a neural network to learn a physics-based model of light transport). For our contribution, 3co is now (1) adapting NeRF-like neural networks to optimally leverage sensor data from various leading devices for 3D computer vision, and (2) forcing these neural networks to learn industry-standard 3D modeling data structures, which can instantly plug-and-play on the leading 3D platforms. As Isaac Newton said, “If I have seen further, it is by standing on the shoulders of giants.” That is, tech giants.
In several ways, TensorFlow is the go-to solution both for prototyping and for large-scale deployment of AI in general. Under-the-hood, TensorFlow uses a sophisticated compiler (XLA) for optimizing how computations are allocated on underlying hardware.