Skip to content
  • Content »
  • Foundations »

Novel view synthesis

View synthesis takes a pictures (or a number of pictures) of a subject and aims to create new views, or so to say novel views, of that subject. 3D reconstruction, the creation of three-dimensional models from a set of images, is a view synthesis that tries to reverse the process of obtaining 2D images from 3D scenes. The task of generating new images that render a subject from a different viewpoint than the one given is called a novel view synthesis.

Approaches to synthesize novel views are for example NeRF and Photogrammetry.

Photogrammetry 3D Scan

...

SDF

Single Distance Functions

Neural Graphics Primitives (NGP)

NeRFs, SDFs, neural images, neural volumes ... are all NGPs

Paper: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
Project Page: https://nvlabs.github.io/instant-ngp/
GitHub: https://github.com/NVlabs/instant-ngp

Installation: Instant-NGP Windows Installation Tutorial

Neural Scene Flow Fields

...

Radiance Fields

...

Plenoxels

Radiance Fields without Neural Networks

https://alexyu.net/plenoxels/

Neural Radiance Fields (NeRFs)

Awesome List: Awesome Neural Radiance Fields - A curated list of awesome neural radiance fields papers
Die Studie: https://www.matthewtancik.com/nerf
gut erklärt: https://www.youtube.com/watch?v=CRlN-cYFxTk
weniger gut erklärt, aber trotzdem: https://youtu.be/WSfEfZ0ilw4
Die Geschichte, Varianten, etc. https://youtu.be/IDMiMKWucaI
Mal angucken: Jon Barron - Understanding and Extending Neural Radiance Fields

novel SOTA in Ultra High Resolution NeRF

SOTA?

4K resolution interactive NeRF rendering support is now possible via work by Alibaba's research team:
https://github.com/frozoul/4K-NeRF

LinkedIn post about it:
https://www.linkedin.com/feed/update/urn:li:activity:7008109018070044674/

Clips in 4K:
https://github.com/frozoul/4K-NeRF/blob/main/4K_results/horns_compare.mp4

The research paper itself:
http://arxiv.org/pdf/2212.04701.pdf

ADOP

Approximate Differentiable One-Pixel Point Rendering, a novel point-based, differentiable neural rendering pipeline.

Video: ADOP: Approximate Differentiable One-Pixel Point Rendering
Video: AI Creates Smooth Videos from Images!
Paper: https://arxiv.org/abs/2110.06635
GitHub: https://github.com/darglein/ADOP

Instant NeRF

Instant NeRF by NVIDIA (see https://developer.nvidia.com/blog/getting-started-with-nvidia-instant-nerfs/, and https://www.nvidia.com/en-us/on-demand/session/siggraph2022-sigg22-s-16/)

Direct Voxel Grid Optimization

Direct Voxel Grid Optimization (short: DVGO; aka DirectVoxGO).
Super-fast Convergence for Radiance Fields Reconstruction.

GitHub: https://github.com/sunset1995/DirectVoxGO
Page: https://sunset1995.github.io/dvgo/
Video: Direct Voxel Grid Optimization: Super-fast Convergence for Radiance Fields Reconstruction

Stereo Radiance Fields (SRF)

Stereo Radiance Fields (SRF) is neural view synthesis approach that is trained end-to-end, generalizes to new scenes in a single forward pass, and requires only sparse views at test time.

Video: Stereo Radiance Fields (SRF): Learning View Synthesis for Sparse Views of Novel Scenes
Page: https://virtualhumans.mpi-inf.mpg.de/srf/
Code: https://github.com/jchibane/srf

TurboNeRF

jperldev:
- https://youtu.be/TeWYAbhgaiU ... video explaining TurboNeRF for Blender
- https://github.com/JamesPerlman/TurboNeRF-Blender
- https://docs.google.com/document/d/1ilywYoQZ9SuD69yE5Pw-fOlJ-lDAAYigG_tIStacP-8/edit ... install guide by Andrew

MegaNeRF

Mega-NeRF:
Scalable Construction of Large-Scale NeRFs for Virtual Fly-Throughs

  • https://github.com/cmusatyalab/mega-nerf
  • https://github.com/cmusatyalab/mega-nerf-viewer

PlenOctrees

Video: PlenOctrees for Real-time Rendering of Neural Radiance Fields
Video: Real-time rendering of NeRFs with PlenOctrees - Angjoo Kanazawa
Page: https://alexyu.net/plenoctrees/
Code: sxyu/volrend PlenOctree Volume Rendering
Code: sxyu/nerfvis (readthedocs()

Darauf basiert wohl der MegaNerf Viewer

Compile sxyu/volrend on M1 Macs

Build instructions are staight forward - only two minor changes needed to build on ARM:

export LIBRARY_PATH=$LIBRARY_PATH:/opt/homebrew/lib/; export ```

CPLUS_INCLUDE_PATH="/usr/local/Cellar/glfw/3.3.4/include"; make -j8

Texture size of the full resolution tree files does exceed some OpenGL limit on Apple silicon ... found some smaller resolution variants (for the web viewer iirc) that work - but I can't recall from where I had them.

NeRF plenoctree viewer on M1 mac.png

NeRF++

Video: Vladlen Koltun: Towards Photorealism (September 2020)

pixelNeRF

Video: pixelNeRF: Neural Radiance Fields from One or Few Images

DONeRF

Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks

Video: https://www.youtube.com/watch?v=6UE1dMUjN_E
Paper: https://arxiv.org/abs/2103.03231
Page: https://depthoraclenerf.github.io

nerfies

https://nerfies.github.io