Core Concepts
Quadric representations enhance accuracy and efficiency in 3D scene modeling for monocular SLAM.
Stats
Neural Radiance Fields (NeRF) have shown promise in monocular SLAM.
NeRFs employ neural networks like MLPs for photo-realistic novel-view synthesis.
3D Gaussian Splatting relies on dense Gaussian representations for effective scene geometries.
Our method exhibits improvements of 6.9% PSNR and 13.2% L1 error compared to other approaches.
Quotes
"Our method involves two key steps: enhancing noisy depth estimations with quadric assumption and concentrating sampling points around quadric planes."
"Quadrics are used as a supervision signal during NeRF network training, improving accuracy."
"Our approach achieves accuracy comparable to methods using ground truth depth."