Neural Radiance Field for Real-time 3D Scene Reconstruction

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Neural Radiance Field for Real-time 3D Scene Reconstruction"

By

KIM Jaehyeok

Abstract:

We present a novel end-to-end framework, NeRFRecon, for 3D scene 
reconstruction that takes a monocular video as its input. Unlike prior 
works predicting the Truncated Signed Distance Function (TSDF) at a fixed 
3D resolution, we propose to utilize Neural Radiance Fields (NeRFs) with a 
3D feature voxel grid for enabling continuous resolution. NeRF regresses 
the radiance (RGB) and density from any arbitrary 3D coordinates with 
trilinear-interpolated feature vectors. Accordingly, the NeRFRecon only 
requires groundtruth RGB-D images for the supervision unlike the baseline 
that requires TSDF and occupancy supervisions. Implicit depth maps and RGB 
images will be composed by querying the optimized NeRF. In addition, 
NeRFRecon performs online optimization in the test time using the lively 
selected keyframes to enhance the specific scene representation. The 
experiments on ScanNet will be illustrated to demonstrate the completeness 
and precision of NeRFRecon compared to the baseline.


Date            : 3 May 2022 (Tuesday)

Time            : 16:25-17:05

Zoom Link:
https://hkust.zoom.us/j/92797640941?pwd=ZWZGcU1aZzR4b1Z4V3dxLytoVTVSUT09

Meeting ID      : 927 9764 0941

Passcode        : csefyp

Advisor         : Dr. XU Dan

2nd Reader      : Dr. CHEN Hao