Hybrid Cost Volume Regularization for Memory-efficient Multi-view Stereo Networks
Qingtian Zhu (Peking University),* Zizhuang Wei (Peking University), Zhongtao Wang (Wuhan University), Yisong Chen (Peking University), GUOPING WANG (PEKING UNIVERSITY)
The 33rd British Machine Vision Conference
Abstract
Learning-based multi-view stereo (MVS) has been studied for years. To overcome the problem of massive computational overhead and memory footprint, different regularization schemes have been attempted. For instance, recurrent methods trade time for space and regularize the cost volume as a sequence, with a RNN to interchange the depth-wise context between sliced cost maps. Meanwhile, cascade methods follow a coarse-to-fine regularization fashion, which enables a gradually refined depth range but still requires a large amount of memory. To this end, we present a novel network for multi-view stereo, termed as HR-MVSNet, which adopts a hybrid design of cascade coarse-to-fine and recurrent cost volume regularization. HR-MVSNet benefits not only from the low memory consumption by the recurrent regularization scheme, but also from the fast inference speed brought by cascade methods. Extensive experiments show that our HR-MVSNet achieves a nice balance between performance and efficiency. It is able to conduct satisfactory reconstruction while still keeps the memory footprint at a relatively low level. For the point clouds and comparative experiments with HR-MVSNet, please contact the first author.Video
Citation
@inproceedings{Zhu_2022_BMVC,
author = {Qingtian Zhu and Zizhuang Wei and Zhongtao Wang and Yisong Chen and GUOPING WANG},
title = {Hybrid Cost Volume Regularization for Memory-efficient Multi-view Stereo Networks},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0073.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by