The network can handle topological variations by modeling a family of shapes in a higher-dimensional space.
Back in June 2021, a team of developers from Google Research had presented a neural network that can handle topological variations by modeling a family of shapes in a higher-dimensional space called HyperNeRF. With this network, the team has managed to bypass the struggle of deformation-based approaches when it comes to model changes in topology.
By lifting NeRFs into a higher-dimensional space, and by representing the 5D radiance field corresponding to each individual input image as a slice through this "hyper-space", the team had achieved more realistic renderings and more accurate geometric reconstructions.
"Our method is inspired by level set methods, which model the evolution of surfaces as slices through a higher dimensional surface. We evaluate our method on two tasks: (I) interpolating smoothly between "moments", i.e., configurations of the scene, seen in the input images while maintaining visual plausibility, and (II) novel-view synthesis at fixed moments," commented the team.
You can learn more about HyperNeRF and its capabilities here. Also, don't forget to join our new Reddit page, our new Telegram channel, follow us on Instagram and Twitter, where we are sharing breakdowns, the latest news, awesome artworks, and more.