Meta tags:

Headings (most frequently used words):

overview, learning, neural, light, fields, with, ray, space, embedding, networks, results, abstract, video, bibtex,

Text of the page (most frequently used words):
and (14), light (12), the (11), #neural (10), fields (10), with (9), ray (9), space (9), results (8), our (8), #embedding (7), network (6), quality (6), field (6), nerf (6), learning (4), radiance (4), are (4), evaluations (4), per (4), rendering (4), method (4), dataset (4), ours (4), networks (3), overview (3), state (3), art (3), view (3), pixel (3), memory (3), that (3), baseline (3), forward (3), facing (3), from (3), all (3), meta (3), benjamin (2), attal (2), jia (2), bin (2), huang (2), michael (2), johannes (2), kopf (2), changil (2), kim (2), video (2), nerfs (2), synthesis (2), however (2), approximate (2), volume (2), integral (2), into (2), data (2), many (2), this (2), paper (2), novel (2), directly (2), integrated (2), along (2), rays (2), evaluation (2), for (2), small (2), datasets (2), only (2), approach (2), stanford (2), addition (2), scenes (2), based (2), speed (2), trade (2), off (2), ground (2), truth (2), more (2), supplemental (2), present (2), inproceedings, attal2022learning, author, zollh, fer, title, booktitle, proceedings, ieee, cvf, conference, computer, vision, pattern, recognition, cvpr, year, 2022, bibtex, produce, they, slow, render, requiring, hundreds, baking, explicit, structures, enables, efficient, but, large, increase, footprint, cases, reduction, propose, representation, contrast, compact, predicts, supports, single, can, also, applied, larger, baselines, few, core, maps, manifold, intermediate, interpolable, latent, achieves, dense, such, sparser, inputs, achieve, competitive, approaches, terms, while, providing, better, far, fewer, abstract, nearest, tarot, horns, 0db, 05s, 4db, 8db, 11s, 1db, 7db, demonstrate, variety, both, sparsely, densely, sampled, show, faithful, reconstruction, highly, complex, dependent, effects, including, comparisons, ablations, found, rendered, images, methods, posted, here, webpage, nex, mpi, shiny, real, challenging, using, popular, coordinate, architectures, leads, poor, mitigate, challenge, leveraging, subdivision, provide, favorable, between, than, previous, shown, graph, below, represents, unlike, which, need, requires, one, code, arxiv, reality, labs, research, zollhöfer, carnegie, mellon, university,

Text of the page (random words):
learning neural light fields with ray space embedding networks learning neural light fields with ray space embedding networks benjamin attal carnegie mellon university jia bin huang meta michael zollhöfer reality labs research johannes kopf meta changil kim meta paper arxiv supplemental results overview video code all results data overview we present neural light fields with ray space embedding a light field directly represents integrated radiance along rays unlike neural radiance fields which need many network evaluations to approximate a volume integral rendering from a light field only requires one evaluation per ray however learning a neural light field is challenging and using popular coordinate based neural network architectures leads to poor view synthesis quality we present a novel ray space embedding approach to mitigate this challenge by leveraging ray space embedding in addition to subdivision our neural light fields provide a more favorable trade off between quality speed and memory than the previous state of the art as shown in the graph below results we demonstrate our method on a variety of scenes both sparsely and densely sampled from the nerf s real forward facing dataset the stanford light fields dataset and nex mpi s shiny dataset our results show faithful reconstruction of highly complex view dependent effects more results including baseline comparisons and ablations are found on our supplemental webpage all rendered images from our method and all baseline methods are posted here ground truth ours 31 7db 2 1s nerf 31 1db 10 4s ground truth ours 36 8db 0 11s nerf 35 4db 13 7s x fields 33 0db 0 05s horns ours nerf tarot small nearest ours nerf x fields abstract neural radiance fields nerfs produce state of the art view synthesis results however they are slow to render requiring hundreds of network evaluations per pixel to approximate a volume rendering integral baking nerfs into explicit data structures enables efficient rendering but results in a large increase in memory footprint and in many cases a quality reduction in this paper we propose a novel neural light field representation that in contrast is compact and directly predicts integrated radiance along rays our method supports rendering with a single network evaluation per pixel for small baseline light field datasets and can also be applied to larger baselines with only a few evaluations per pixel at the core of our approach is a ray space embedding network that maps the 4d ray space manifold into an intermediate interpolable latent space our method achieves state of the art quality on dense forward facing datasets such as the stanford light field dataset in addition for forward facing scenes with sparser inputs we achieve results that are competitive with nerf based approaches in terms of quality while providing a better speed quality memory trade off with far fewer network evaluations overview video bibtex inproceedings attal2022learning author benjamin attal and jia bin huang and michael zollh o fer and johannes kopf and changil kim title learning neural light fields with ray space embedding networks booktitle proceedings of the ieee cvf conference on computer vision and pattern recognition cvpr year 2022