RoomNeRF: Representing Empty Room as Neural Radiance Fields for View Synthesis
Mangyu Kong (Yonsei University),* Seongwon Lee (Yonsei university), Euntai Kim (Yonsei University)
The 34th British Machine Vision Conference
Abstract
We present a method for novel view synthesis of empty rooms from object-existing room images. Despite the remarkable achievements of previous inpainted NeRFs for object removal tasks, they have a limitation in completely reconstructing the empty room due to the lack of consideration for the room characteristic. Our proposed network, named RoomNeRF, is designed to fully exploit the shared intrinsic properties of each plane of the room via the Pattern Transfer (PT) and Planar Constraint (PC). For each plane, the PT and PC modules capture shared visual patterns and geometrical structures, respectively, and transfer them to areas occluded by objects, enabling realistic empty room reconstructions without being disturbed by invisible areas of the input images. With these internal learning strategies, RoomNeRF successfully performs novel view synthesis of an empty room from multi-object presence images in extensive experiments and demonstrates its superiority.Video
Citation
@inproceedings{Kong_2023_BMVC,
author = {Mangyu Kong and Seongwon Lee and Euntai Kim},
title = {RoomNeRF: Representing Empty Room as Neural Radiance Fields for View Synthesis},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0825.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by