Videos with six degrees of freedom (6-DoF) let viewers freely explore an area by allowing them to adjust their head position (3 degrees of freedom) and direction (3 degrees of freedom). As a result, 6-DoF movies provide immersive experiences with various fascinating AR/VR applications. View synthesis, which produces new, unseen perspectives of an environment—static or dynamic—from a sequence of posed photos or videos, is the basic mechanism that powers 6-DoF video. Recent advances have been made in photorealistic view synthesis for static situations using volumetric scene representations like neural radiance fields and immediate neural graphics primitives.
It remains difficult to develop a 6-DoF video format that can achieve good quality, quick rendering, and a compact memory footprint, despite various recent research building dynamic view synthesis pipelines on top of these volumetric representations (even given many synchronized video streams from multi-view camera rigs ). Generating a single-megapixel image using current methods for memory-efficient 6-DoF video can take close to a minute. Even for brief video clips, works that aim to render quickly and immediately portray dynamic volumes with 3D textures need terabytes of storage. While prior volumetric methods use compressed or sparse volume storage to optimize memory use and performance for static scenes, only recent work has addressed applying similar strategies to dynamic scenes.
None of the abovementioned depictions are very good at capturing highly view-dependent appearances, including reflections and refractions brought on by non-planar surfaces. In this study, they introduce HyperReel, a unique 6-DoF video representation that provides state-of-the-art quality while being memory-efficient and real-time renderable at high resolution. A unique rayconditioned sample prediction network, which forecasts sparse point samples for volume rendering, is the initial component of their methodology. Their solution is distinctive in that it (1) speeds up volume rendering and (2) enhances rendering quality for difficult view-dependent situations, in contrast to previous static view synthesis methods that employ sample networks.
Second, using the spatiotemporal redundancy of a dynamic scene, they provide a memory-efficient dynamic volume representation that achieves a high compression rate. To more specifically express a series of volumetric keyframes compactly, they enhance Tensorial Radiance Fields, and they capture intermediate frames with trainable scene flow. Their high-fidelity 6-DoF video representation, HyperReel, is made up of a mixture of these two methods. With comparisons to cutting-edge sampling network-based techniques for static scenes and 6-DoF video representations for dynamic scenes, they verify the separate components of their methodology and their representation as a whole.
HyperReel not only performs better than these earlier efforts, but it also offers stunning representations for difficult non-Lambertian looks. Without utilizing any customized CUDA code, their system renders up to 18 frames per second at a megapixel resolution. In conclusion, the contributions made by their effort are as follows: 1. A brand-new volumetric view synthesis sample prediction network that quickens volume rendering while effectively representing intricate view-dependent impacts 2. A dynamic scene is compactly represented using a memory-efficient dynamic volume representation. 3. HyperReel, a 6-DoF video representation that balances speed, quality, and memory while rendering in real-time at megapixel resolution.
Check out the Paper, Code, and Project. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our Reddit Page, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.
Aneesh Tickoo is a consulting intern at MarktechPost. He is currently pursuing his undergraduate degree in Data Science and Artificial Intelligence from the Indian Institute of Technology(IIT), Bhilai. He spends most of his time working on projects aimed at harnessing the power of machine learning. His research interest is image processing and is passionate about building solutions around it. He loves to connect with people and collaborate on interesting projects.