README.md

metadata

license: mit
pretty_name: Video2BEV
size_categories:
  - 10B<n<100B

Video2BEV-Open

Yet another distribution of the dataset from the paper: 📜 Video2BEV

Package and Data Structure

Package

The dataset is archived using Zstandard compression, e.g., train-shard00.tar.zst or test-shard01.tar.zst.

• Extract with the command

tar xaf train-shard00.tar.zst

• Extract all

for file in {train,test}-shard*.tar.zst; do
  tar xaf "$file"
done

Data Structure

A dataset entry looks like:

# /train
0000
├── drone
│   ├── angle-30.mp4
│   └── angle-45.mp4
├── google
│   ├── 1.Studio_Art%2C_Reed_College.jpg
│   ├── 12.740x600_Reed_02.jpg
│   ├── 17.ReedCollegePAB1.jpg
│   ├── 2.art-buidling.jpg
│   ├── 21.etcse3s6r1sk5skh.jpg
│   └── 4.Studio_Art_Building%2C_Reed_College%2C_September_2013.jpg
├── satellite
│   └── 1235.jpg
└── street
    ├── 1.jpg
    ├── 2.jpg
    └── 3.jpg

# /test
0000
├── drone
│   ├── angle-30.mp4
│   └── angle-45.mp4
├── satellite
│   └── 0010.jpg
└── street
    ├── 1.jpg
    ├── 2.jpg
    ├── 3.jpg
    ├── 4.jpg
    └── 5.jpg

Guarantees

• ID All IDs are zero-padded 4-digit strings (%04d) and fall in the range 0000-1651.
• Drone There are exactly two videos, angle-30.mp4 and angle-45.mp4, under ${split}/${id}/drone/. (H.264, 10 FPS, 512×512)
• Satellite There is exactly one JPEG file under ${split}/${id}/satellite/${id}.jpg.
• Google Only the training split ships with google/ folders, and not every ID has one.

Exceptions

• test/0274/street/ contains some images that appear to belong in a google/ folder and do not follow the usual naming convention. This may be due to an issue in the original packaging. Since this repository is a reorganization and redistribution (not the original release), the files are provided as-is.
• In the test split, 158 IDs have no street/ folder.