Skip to content


Switch branches/tags



This package provides a minimal set of tools for working with the KITTI dataset [1] in Python. So far only the raw datasets and odometry benchmark datasets are supported, but we're working on adding support for the others. We welcome contributions from the community.


Using pip

You can install pykitti via pip using

pip install pykitti

From source

To install the package from source, simply clone or download the repository to your machine

git clone

and run the provided setup tool

cd pykitti
python install


This package assumes that you have also downloaded the calibration data associated with the sequences you want to work on (these are separate files from the sequences themselves), and that the directory structure is unchanged from the original structure laid out in the KITTI zip files.


Homogeneous coordinate transformations are provided as 4x4 numpy.array objects and are denoted as T_destinationFrame_originFrame.

Pinhole camera intrinsics for camera N are provided as 3x3 numpy.array objects and are denoted as K_camN. Stereo pair baselines are given in meters as b_gray for the monochrome stereo pair (cam0 and cam1), and b_rgb for the color stereo pair (cam2 and cam3).


More detailed examples can be found in the demos directory, but the general idea is to specify what dataset you want to load, then access the parts you need and do something with them.

Camera and velodyne data are available via generators for easy sequential access (e.g., for visual odometry), and by indexed getter methods for random access (e.g., for deep learning). Images are loaded as PIL.Image objects using Pillow.

import pykitti

basedir = '/your/dataset/dir'
date = '2011_09_26'
drive = '0019'

# The 'frames' argument is optional - default: None, which loads the whole dataset.
# Calibration, timestamps, and IMU data are read automatically. 
# Camera and velodyne data are available via properties that create generators
# when accessed, or through getter methods that provide random access.
data = pykitti.raw(basedir, date, drive, frames=range(0, 50, 5))

# dataset.calib:         Calibration data are accessible as a named tuple
# dataset.timestamps:    Timestamps are parsed into a list of datetime objects
# dataset.oxts:          List of OXTS packets and 6-dof poses as named tuples
# dataset.camN:          Returns a generator that loads individual images from camera N
# dataset.get_camN(idx): Returns the image from camera N at idx  
# dataset.gray:          Returns a generator that loads monochrome stereo pairs (cam0, cam1)
# dataset.get_gray(idx): Returns the monochrome stereo pair at idx  
# dataset.rgb:           Returns a generator that loads RGB stereo pairs (cam2, cam3)
# dataset.get_rgb(idx):  Returns the RGB stereo pair at idx  
# dataset.velo:          Returns a generator that loads velodyne scans as [x,y,z,reflectance]
# dataset.get_velo(idx): Returns the velodyne scan at idx  

point_velo = np.array([0,0,0,1])
point_cam0 =

point_imu = np.array([0,0,0,1])
point_w = [ for o in data.oxts]

for cam0_image in data.cam0:
    # do something

cam2_image, cam3_image = data.get_rgb(3)


PIL Image data can be converted to an OpenCV-friendly format using numpy and cv2.cvtColor:

img_np = np.array(img)
img_cv2 = cv2.cvtColor(img_np, cv2.COLOR_RGB2BGR)

Note: This package does not actually require that OpenCV be installed on your system, except to run


[1] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, "Vision meets robotics: The KITTI dataset," Int. J. Robot. Research (IJRR), vol. 32, no. 11, pp. 1231–1237, Sep. 2013. `