ONCE

Downloads

We maintain the data using Google Drive for global users and Baidu Yunpan for Chinese users. Extraction code for all Baidu Yunpan download is 1234

Download link for annotations: [Baidu Yunpan]

Download link for cameras: [Baidu Yunpan]

Download link for lidars: [Baidu Yunpan]

Training split

annotations: [google drive] lidar data: [google drive] camera data: [google drive]

Validation split

annotations: [google drive] lidar data: [google drive] camera data: [google drive]

Testing split

annotations: [google drive] lidar data: [google drive] camera data: [google drive]

Unlabeled small split

annotations: [google drive] lidar data: [google drive] camera data: [google drive]

Unlabeled medium split

annotations: [google drive]lidar data p1: [google drive]lidar data p2: [google drive]

lidar data p3: [google drive] lidar data p4: [google drive]camera data p1: [google drive]

camera data p2: [google drive] camera data p3: [google drive] camera data p4: [google drive]

Unlabeled large split

annotations: [google drive]lidar data p5: [google drive]lidar data p6: [google drive]

lidar data p7: [google drive] lidar data p8: [google drive]lidar data p9: [google drive]

camera data p5: [google drive] camera data p6: [google drive] camera data p7: [google drive]

camera data p8: [google drive] camera data p9: [google drive]

Data orgainization

We organize the dataset by sensors and split unlabeled data into three different scales. Users can conviently choose which part of data to download according to different training intensions. Tar files larger than 20G are split into multiple files by linux split command with parta* extensions.

data_root
└─3D_DATA
    ├─3D_infos
    │      train_infos.tar
    │      val_infos.tar
    │      test_infos.tar
    │      raw_small_infos.tar
    │      raw_medium_infos.tar
    │      raw_large_infos.tar
    ├─3D_images
    │      train_cam0[1356789].tar
    │      val_cam0[1356789].tar
    │      test_cam0[1356789].tar
    │      raw_cam01_p[0-9].tar
    │      raw_cam03_p[0-9].tar
    │      raw_cam05_p[0-9].tar
    │      raw_cam06_p[0-9].tar
    │      raw_cam07_p[0-9].tar
    │      raw_cam08_p[0-9].tar
    │      raw_cam09_p[0-9].tar
    └─3D_lidars
           train_lidar.tar
           val_lidar.tar
           test_lidar.tar
           raw_lidar_p0.tar.parta*
           raw_lidar_p1.tar.parta*
           raw_lidar_p2.tar.parta*
           raw_lidar_p3.tar.parta*
           raw_lidar_p4.tar.parta*
           raw_lidar_p5.tar.parta*
           raw_lidar_p6.tar.parta*
           raw_lidar_p7.tar.parta*
           raw_lidar_p8.tar.parta*
           raw_lidar_p9.tar.parta*

After downloading desired part of data, user should put all downloaded files into the SAME folder discarding the formor folder structure. For example, both train_infos.tar, train_lidar.tar and train_cam0[1356789].tar should put in the same folder. After concatenation and extraction, file structure under data root directory is something like this

data_root
└─data
    ├─000000
    │   ├─ cam01
    │   │    ├─ frame_timestamp_1.jpg
    │   │    ├─ frame_timestamp_2.jpg
    │   │    ├─ ....
    │   │    └─ frame_timestamp_n.jpg
    │   ├─ cam03
    │   │    ├─ ...
    │   ├─ cam05
    │   │    ├─ ...
    │   ├─ cam06
    │   │    ├─ ...
    │   ├─ cam07
    │   │    ├─ ...
    │   ├─ cam08
    │   │    ├─ ...
    │   ├─ cam09
    │   │    ├─ ...
    │   ├─ lidar_roof
    │   │    ├─ frame_timestamp_1.bin
    │   │    ├─ frame_timestamp_2.bin
    │   │    ├─ ....
    │   │    └─ frame_timestamp_n.bin
    │   └─ 000000.json
    ├─000001
    ├─000002
    └─...

where n corresponds to the number of scenes in the specific road.