KITTI-360

SSCBench-KITTI-360

Change Log

• 2024/09: SSCBench-KITTI-360 is now available on Hugging Face
• 2023/06: Initial release.

Overview

KITTI-360 covers a driving distance of 73.7km, comprising 300K images and 80K laser scans. Leveraging the open-source training and validation set, we build SSCBench-KITTI-360 consisting of 9 long sequences. To reduce redundancy, we sample every 5 frames following the SemanticKITTI SSC benchmark. The training set includes 8,487 frames from scenes 00, 02-05, 07, and 10, while the validation set comprises 1,812 frames from scene 06. The testing set comprises 2,566 frames from scene 09. In total, the dataset contains 12,865 frames

License

The KITTI-360 dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License ("CC BY-NC-SA 3.0"). Please check their website for more details.

We release SSCBench-KITTI-360 under the same license. When you download or use the SSCBench-KITTI-360 from the website or elsewhere, you are agreeing to comply with the terms of the "CC BY-NC-SA 3.0" License.

Folder Structure and format

The folder structure of the SSCBench-KITTI-360 dataset is as follows:

dataset/kitti360/
|-- KITTI-360 
|   |-- data_2d_raw
|   |   |-- 2013_05_28_drive_0000_sync # train:[0, 2, 3, 4, 5, 7, 10] + val:[6] + test:[9]
|   |   |   |-- image_00
|   |   |   |   |-- data_rect # RGB images for left camera
|   |   |   |   |   |-- 000000.png
|   |   |   |   |   |-- 000001.png
|   |   |   |   |   |-- ...
|   |   |   |   |-- timestamps.txt
|   |   |   |-- image_01
|   |   |   |   |-- data_rect # RGB images for right camera
|   |   |   |   |   |-- 000000.png
|   |   |   |   |   |-- 000001.png
|   |   |   |   |   |-- ...
|   |   |   |   |-- timestamps.txt
|   |   |   |-- voxels # voxelized point clouds
|   |   |   |   |-- 000000.bin # voxelized input
|   |   |   |   |-- 000000.invalid # voxelized invalid mask
|   |   |   |   |-- 000000.label  #voxelized label
|   |   |   |   |-- 000005.bin # calculate every 5 frames 
|   |   |   |   |-- 000005.invalid
|   |   |   |   |-- 000005.label
|   |   |   |   |-- ...
|   |   |   |-- cam0_to_world.txt
|   |   |   |-- pose.txt # car pose information
|   |   |-- ...
|   |   |-- 2013_05_28_drive_0010_sync 
|   |-- preprocess # preprocessed downsampled labels
|   |   |-- labels # not unified
|   |   |   |-- 2013_05_28_drive_0000_sync 
|   |   |   |   |-- 000000_1_1.npy # original labels
|   |   |   |   |-- 000000_1_8.npy # 8x downsampled labels
|   |   |   |   |-- 000005_1_1.npy
|   |   |   |   |-- 000005_1_8.npy
|   |   |   |   |-- ...
|   |   |   |-- ... 
|   |   |   |-- 2013_05_28_drive_0010_sync
|   |   |-- labels_half # not unified, downsampled 
|   |   |   |-- 2013_05_28_drive_0000_sync 
|   |   |   |   |-- 000000_1_1.npy # original labels
|   |   |   |   |-- 000000_1_8.npy # 8x downsampled labels
|   |   |   |   |-- 000005_1_1.npy
|   |   |   |   |-- 000005_1_8.npy
|   |   |   |   |-- ...
|   |   |   |-- ... 
|   |   |   |-- 2013_05_28_drive_0010_sync
|   |   |-- unified # unified
|   |   |   |-- labels
|   |   |   |   |-- 2013_05_28_drive_0000_sync 
|   |   |   |   |   |-- 000000_1_1.npy # original labels
|   |   |   |   |   |-- 000000_1_8.npy # 8x downsampled labels
|   |   |   |   |   |-- 000005_1_1.npy
|   |   |   |   |   |-- 000005_1_8.npy
|   |   |   |   |   |-- ...
|   |   |   |   |-- ... 
|   |   |   |   |-- 2013_05_28_drive_0010_sync
|   |-- calibration # preprocessed downsampled labels
|   |   |-- calib_cam_to_pose.txt
|   |   |-- calib_cam_to_velo.txt
|   |   |-- calib_sick_to_velo.txt
|   |   |-- image_02.yaml
|   |   |-- image_03.yaml
|   |   |-- perspective.txt

For each frame in the dataset, we provide the following information:

• image_00: RGB image of size 376x1408.
• voxels: Voxelized point cloud of size 256x256x32. Each voxel is a 0.2x0.2x0.2 meter cube.
  • *.bin: voxelized point cloud in binary format
  • *.label: voxelized point cloud label in binary format
  • *.invalid = voxelized point cloud invalid mask in binary format.

For MonoScene and VoxFormer, a preprocessed downsampled version of the dataset is provided in the preprocess folder. We provide two scales of downsampled point clouds: 1/1 and 1/8. The downsampled point clouds are stored in the labels folder, stored as .npy files.

Data Download

The dataset can be downloaded from Hugging Face. The dataset is provided in the form of squashed file system for easy use for singularity containers. Use the following command to merge the split parts of the datasets.

cat split_parts_* > combined.sqfs

• If you want to use the dataset on a singularity container, you can mount each squashed file system to the container using the --overlay option.
• If you want to use the dataset on a normal system, you can unsquash the file system using the unsquashfs command (more details here). Then, please organize the data as the folder structure described above.