Written and maintained by CASRAI Editorial Board
Last updated
Short answer: the KITTI dataset is a public collection of autonomous-driving sensor recordings and benchmarks from the Karlsruhe Institute of Technology and the Toyota Technological Institute at Chicago, built from a camera-, laser-scanner- and GPS-equipped car driven around Karlsruhe, Germany. It is released under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 licence, and each benchmark asks to be cited.
Definition and scope
KITTI (also written “KITTI Vision Benchmark Suite”) is a project of Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago, created by A. Geiger, P. Lenz, C. Stiller and R. Urtasun. The project site went online on 20 March 2012 with stereo, optical flow and visual odometry benchmarks, and later added 3D object detection, tracking, road, depth, semantic and scene flow benchmarks and multi-object tracking and segmentation (MOTS).
The name is a proper noun for a dataset. It is unrelated to kitting, the logistics practice of assembling items into kits, and to the personal name Kitti.
What the recording platform captured
The recordings come from a VW Passat station wagon driven in the mid-size city of Karlsruhe, in rural areas and on highways. The 2013 International Journal of Robotics Research (IJRR) paper “Vision meets Robotics: The KITTI Dataset” describes about six hours of traffic scenarios recorded at 10 to 100 Hz. The sensors and data formats it lists:
| Component | As described in the IJRR paper |
|---|---|
| Cameras | Two grayscale (PointGrey Flea2 FL2-14S3M-C) and two colour (FL2-14S3C-C), 1.4 megapixel, global shutter; stereo baseline about 54 cm |
| Laser scanner | Velodyne HDL-64E: 64 beams, 360 degree horizontal field, 10 Hz, range about 120 m, about 1.3 million points per second |
| Localisation | OXTS RT3003 GPS/IMU at 100 Hz |
| Image files | 8-bit PNG, lossless |
| Laser files | Float binaries (x, y, z, reflectance), about 120,000 points per frame |
| GPS/IMU files | 30 values per frame |
| Annotations | 3D bounding-box tracklets; classes Car, Van, Truck, Pedestrian, Person (sitting), Cyclist, Tram, Misc |
The paper also notes that the Velodyne scanner has a rolling shutter, so three timestamp files are provided, and that GPS/IMU outages shorter than one second were interpolated, with the last three entries set to -1. Anyone processing the raw files should read these details before synchronising sensors.
Benchmarks and related concepts
The benchmarks were introduced in “Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite” (Geiger, Lenz and Urtasun, CVPR 2012). Each benchmark has its own page on the project site and its own citation request, so a paper using the 3D object detection benchmark cites differently from one using only the raw data. Related reading on this site: image annotation (how labelled boxes like the tracklets are produced) and remote sensing (the wider family of sensing methods that includes lidar).
Where it matters: citation and licence checklist
- Licence: the project site states the data is published under Creative Commons Attribution-NonCommercial-ShareAlike 3.0. NonCommercial terms matter for industry-funded or commercialised work; see Creative Commons and Creative Commons licences for research data. Interpretation for a specific use is for your institution’s research office or counsel.
- Citation: the site asks users to cite the work when using the data, with the relevant paper per benchmark family. Recording the benchmark name and the retrieval date supports data citation.
- Scope of the raw release: the IJRR paper states the raw data contains about 25% of the overall recordings, mainly sequences with 3D tracklet annotations, with benchmark test sequences removed.
- Conditions of capture: daytime recordings from 26, 28, 29 and 30 September and 3 October 2011, in and around Karlsruhe, in five categories: Road, City, Residential, Campus and Person. Results may not transfer to other countries, seasons, lighting or sensor setups.
Common misunderstandings
- “KITTI is one benchmark.” It is a suite of benchmarks built on shared recordings.
- “The raw download is everything recorded.” By the authors’ account it is about a quarter of the recordings.
- “Open means unrestricted.” The data is openly available but under a NonCommercial, ShareAlike licence.
- “The sensors are current.” The hardware is a 2011-era configuration, as the recording dates show.
FAQ
Who created the KITTI dataset?
A. Geiger, P. Lenz, C. Stiller and R. Urtasun, as a project of Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago.
Can KITTI be used commercially?
The published licence is Creative Commons Attribution-NonCommercial-ShareAlike 3.0, which restricts commercial use. Confirm any specific use with your institution.
How big is the raw data?
The IJRR paper gives 180 GB for the provided data.
Sources
- KITTI Vision Benchmark Suite, cvlibs.net/datasets/kitti, read 2026-11-08.
- Geiger, Lenz, Stiller, Urtasun (2013), “Vision meets Robotics: The KITTI Dataset”, International Journal of Robotics Research, PDF, read 2026-11-08.
- Geiger, Lenz, Urtasun (2012), “Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite”, CVPR, PDF.
