本文介绍CasMVSNet的复现情况。该网络用于多视图立体匹配任务,可估计场景密集深度以重建三维场景。文中涉及训练数据集解压、训练、测试数据集处理、测试及结果评估等步骤,复现的定量结果与原文接近,整体指标0.357,略高于原文的0.355,表明复现成功。
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜
![[cvpr 2020]casmvsnet_paddle复现 - 创想鸟](https://www.chuangxiangniao.com/wp-content/themes/justnews/themer/assets/images/lazy.png)
1. CasMVSNet简介
来源:Cascade Cost Volume for High-Resolution Multi-View Stereo and Stereo Matching
项目意义:Mult-View Stereo(MVS)任务中旨在根据多视图输入估计场景的密集深度表示,从而重建三维场景。
2. 参考资料
CasMVSNet论文:Cascade Cost Volume for High-Resolution Multi-View Stereo and Stereo MatchingPytorch代码:alibaba/cascade-stereo
3. 训练代码
3.1 解压训练数据集
In [2]
!rar x data/data129222/dtu_training.rar data/!rm -rf data/data129222/dtu_training.rar !unzip -d data/mvs_training/dtu/Depths_raw data/data129777/Depths_raw1.zip!rm -rf data/data129777/Depths_raw1.zip !unzip -d data/mvs_training/dtu/Depths_raw data/data129777/Depths_raw2.zip!rm -rf data/data129777/Depths_raw2.zip '''rar x data/data129222/dtu_training.rar data/rm -rf data/data129222/dtu_training.rar unzip -d data/mvs_training/dtu/Depths_raw data/data129777/Depths_raw1.ziprm -rf data/data129777/Depths_raw1.zip unzip -d data/mvs_training/dtu/Depths_raw data/data129777/Depths_raw2.ziprm -rf data/data129777/Depths_raw2.zip '''
RAR 5.30 beta 2 Copyright (c) 1993-2015 Alexander Roshal 4 Aug 2015Trial version Type RAR -? for helpCannot open data/data129222/dtu_training.rar没有那个文件或目录No files to extractunzip: cannot find or open data/data129777/Depths_raw1.zip, data/data129777/Depths_raw1.zip.zip or data/data129777/Depths_raw1.zip.ZIP.unzip: cannot find or open data/data129777/Depths_raw2.zip, data/data129777/Depths_raw2.zip.zip or data/data129777/Depths_raw2.zip.ZIP.
'nrar x data/data129222/dtu_training.rar data/nrm -rf data/data129222/dtu_training.rar nunzip -d data/mvs_training/dtu/Depths_raw data/data129777/Depths_raw1.zipnrm -rf data/data129777/Depths_raw1.zip nunzip -d data/mvs_training/dtu/Depths_raw data/data129777/Depths_raw2.zipnrm -rf data/data129777/Depths_raw2.zip n'
3.2 开始训练
In [ ]
cd /home/aistudio/work/CasMVSNet_paddle/!python train.py --dataset=dtu_yao --trainpath /home/aistudio/data/mvs_training/dtu/ --testpath /home/aistudio/data/mvs_training/dtu/ --trainlist lists/dtu/train.txt --testlist lists/dtu/test.txt --batch_size 2
4. 测试代码
4.1 解压测试数据集
In [ ]
!unzip -d data/ data/data129222/dtu_testing.zip
4.2 安装测试所需的包
In [ ]
!pip install plyfile
4.3 开始测试
In [ ]
cd /home/aistudio/work/CasMVSNet_paddle/!python test.py --dataset=general_eval --batch_size=1 --testpath /home/aistudio/data/dtu_testing --testlist lists/dtu/test.txt --loadckpt ./log/model_000015.ckpt
5. 结果评估
由于测试需要MATLAB,所以需要将测试生成的PLY点云文件下载到本地,然后在本地评估。评估的数据为DTU数据集,评估代码和数据:http://roboimagedata.compute.dtu.dk/?page_id=36
6. 结果展示
定量结果表明,复现成功。overall指标几乎没有差别。
复现结果: acc.: 0.369 comp.:0.346 overall:0.357
原文结果: acc.: 0.325 comp.:0.385 overall:0.355
部分可视化结果展示:![[CVPR 2020]CasMVSNet_Paddle复现 - 创想鸟](https://www.chuangxiangniao.com/wp-content/themes/justnews/themer/assets/images/lazy.png)
以上就是[CVPR 2020]CasMVSNet_Paddle复现的详细内容,更多请关注创想鸟其它相关文章!
版权声明:本文内容由互联网用户自发贡献,该文观点仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。
如发现本站有涉嫌抄袭侵权/违法违规的内容, 请发送邮件至 chuangxiangniao@163.com 举报,一经查实,本站将立刻删除。
发布者:程序猿,转转请注明出处:https://www.chuangxiangniao.com/p/58024.html
微信扫一扫
支付宝扫一扫