The Path to Nerf

LLFF

• Prepare Image

data/desk/images
├── 2c68275866dc2b1da47f5eaf885c583.jpg
├── 2fd70dcc8b5f374886df488054fa5ab.jpg
├── 34df512a279040a3820c531c13d5e17.jpg
├── 358af944c88b7a3ff231e4e2eeafdd0.jpg
├── 3808e607c1cf63cebf12b29dc7f2442.jpg
├── 5dbf835776758757c2be8b5ee08e17b.jpg
├── 9ca4a35fffb7a69f84be9e1f2945848.jpg
├── d86ac4a847edb12b6dcded776d47732.jpg
├── e1402fe72717cfac86bd9cd69c3b98f.jpg
├── f2dfe38840ae9fd02dd33fa09dad769.jpg
├── fd177d4977813dfce11b09f71797e01.jpg
└── ff21b26780837071d7980447866f217.jpg

• At LLFF project directory

nvidia-docker run --rm --env CUDA_VISIBLE_DEVICES=3 --volume /:/host --workdir /host$PWD tf_colmap bash demo.sh

• --env CUDA_VISIBLE_DEVICES=3 is docker container env.
• Use watch command to check which gpu is available
  • watch 'nvidia-smi | grep Default'
  • 
• output

data/desk
├── colmap_output.txt
├── database.db
├── images
├── images_480x360
├── mpis_360
├── outputs
├── poses_bounds.npy
└── sparse
    └── 0
        ├── cameras.bin
        ├── images.bin
        ├── points3D.bin
        └── project.ini

Attention

• 拍摄的图片需要是个矩阵，不然容易报错
  • 报错代码主要位置
  • llff/poses/pose_utils.py:66

Docker 越权

• docker run -it --volume /:/host tf_colmap
• change /host/etc/sudoers
• append lihuibin ALL=(ALL:ALL) ALL below root ALL=(ALL:ALL) ALL
• linux下添加用户并赋予root权限

Nerf

• issues
  • ffmpeg: Unrecognized option 'crf'. Error splitting the argument list: Option not found
    • stackoverflow
    • remove ffmpeg from conda
      • conda remove --force ffmpeg
    • install ffmpeg with apt
      • sudo apt install ffmpeg