TimeNeRF: Building Generalizable Neural Radiance Fields across Time from Few-Shot Input Views
Authors:
- Hsiang-Hui HungHHHsiang-Hui Hung
- Huu-Phu DoHDHuu-Phu Do
- Yung-Hui LiYLYung-Hui Li
- Ching-Chun HuangCHChing-Chun Huang
Preprint
Abstract
We present TimeNeRF, a generalizable neural rendering approach for rendering novel views at arbitrary viewpoints and at arbitrary times, even with few input views. For real-world applications, it is expensive to collect multiple views and inefficient to re-optimize for unseen scenes. Moreover, as the digital realm, particularly the metaverse, strives for increasingly immersive experiences, the ability to model 3D environments that naturally transition between day and night becomes paramount. While current techniques based on Neural Radiance Fields (NeRF) have shown remarkable proficiency in synthesizing novel views, the exploration of NeRF's potential for temporal 3D scene modeling remains limited, with no dedicated datasets available for this purpose. To this end, our approach harnesses the strengths of multi-view stereo, neural radiance fields, and disentanglement strategies across diverse datasets. This equips our model with the capability for generalizability in a few-shot setting, allows us to construct an implicit content radiance field for scene representation, and further enables the building of neural radiance fields at any arbitrary time. Finally, we synthesize novel views of that time via volume rendering. Experiments show that TimeNeRF can render novel views in a few-shot setting without per-scene optimization. Most notably, it excels in creating realistic novel views that transition smoothly across different times, adeptly capturing intricate natural scene changes from dawn to dusk.Keywords
All Related Versions
This publication has 27 references indexed in Scilit:
- AdaNeRF: Adaptive Sampling for Real-Time Rendering of Neural Radiance Fields2022
- Direct Voxel Grid Optimization: Super-fast Convergence for Radiance Fields Reconstruction2022
- TermiNeRF: Ray Termination Prediction for Efficient Neural Rendering2021
- IBRNet: Learning Multi-View Image-Based Rendering2021
- D-NeRF: Neural Radiance Fields for Dynamic Scenes2021
- CoMoGAN: continuous model-guided image-to-image translation2021
- TSIT: A Simple and Versatile Framework for Image-to-Image Translation2020
- Deep Network Interpolation for Continuous Imagery Effect Transition2019
- DLOW: Domain Flow for Adaptation and Generalization2019
- The Unreasonable Effectiveness of Deep Features as a Perceptual Metric2018