Related Experiment Video
Updated: Jun 8, 2025
14:25
Determining 3D Flow Fields via Multi-camera Light Field Imaging
Published on: March 6, 2013
16.6K
Neural Radiance Fields for Fisheye Driving Scenes Using Edge-Aware Integrated Depth Supervision
1Division of Electronics and Information Engineering, Jeonbuk National University, Jeonju 54896, Republic of Korea.
Sensors (Basel, Switzerland)
|November 9, 2024
Summary
This study introduces an edge-aware loss function for Neural Radiance Fields (NeRF) to generate realistic driving scene views from fisheye camera images. The method effectively handles distortions and improves novel view synthesis using LiDAR and depth data.
Area of Science:
- Computer Vision
- Computer Graphics
- Robotics
Background:
- Neural Radiance Fields (NeRF) excel at novel view synthesis but are limited by pinhole camera assumptions.
- Driving scenarios present unique challenges due to fisheye cameras' wide field of view and image distortion.
Purpose of the Study:
- To adapt NeRF for driving scenarios captured with fisheye cameras.
- To develop an effective method for synthesizing photorealistic novel views from distorted, wide-angle imagery.
Main Methods:
- Proposed an edge-aware integration loss function for NeRF.
- Leveraged sparse LiDAR projections and learning-based dense depth maps.
- Assigned greater weights to points with depth values similar to sensor data.
Main Results:
- Demonstrated effectiveness on KITTI-360 and JBNU-Depth360 datasets.
- Achieved superior performance in novel view synthesis compared to existing methods.
- Successfully synthesized photorealistic images from fisheye driving data.
Conclusions:
- The proposed edge-aware NeRF approach overcomes limitations of traditional methods for fisheye camera data.
- This technique significantly enhances the synthesis of novel views in complex driving environments.
- The method shows strong potential for applications in autonomous driving and scene reconstruction.
