GeoNeRF: Generalizing NeRF with Geometry Priors

Advances in Neural Rendering

“…Joint training on large datasets enable training without multi‐view supervision. GeoNeRF [JLF21] constructs a set of cascaded cost volumes and employs transformers to infer geometry and appearance.…”

RefiNeRF: Modelling dynamic neural radiance fields with inconsistent or missing camera parameters

“…Significant progress has been made in recent years to model complex geometries using methods such as point-clouds [11,3,39,32], voxels [30,47,40,16], octrees [38,45], or various computed tomography algorithms [6]. These computationally expensive techniques have served as a bottleneck for true 3D understanding and most tasks require strong priors [7,28,14] or existing templates [25,10,42,41,37]. In NeRFs [23] implicit functions are used for representing scenes by implicitly encoding photometric attributes such as colour, surface illumination, opacity, etc., using shallow neural nets [21,1,26,22].…”

wildNeRF: Novel view synthesis of in-the-wild dynamic scenes using using sparse monocular view data

“…using shallow neural nets [3][4][5]35], an important characteristic of these approaches is that they are usually selfsupervised. The models can be trained in an end-to-end manner by leveraging a pixel-wise photometric reconstruction loss, which is in strong contrast to existing approaches that require strong priors [36][37][38] or existing templates [2,[39][40][41][42], each of which are challenging to acquire and significantly limit the generalizability of these approaches to ideal pre-designed scenes.…”