KITTI-CARLA: a KITTI-like dataset generated by CARLA Simulator

TL;DR

KITTI-CARLA is a synthetic dataset generated from the CARLA simulator, designed to facilitate transfer learning and benchmarking for autonomous vehicle perception tasks by mimicking the KITTI dataset setup.

Contribution

This paper introduces a new synthetic dataset that replicates the KITTI sensor setup using CARLA, enabling improved transfer learning and benchmarking in autonomous driving research.

Findings

Dataset includes 7 sequences with 5000 frames each.

Provides diverse environments: city, suburban, mountain, rural, highway.

Facilitates transfer learning from synthetic to real data.

Abstract

KITTI-CARLA is a dataset built from the CARLA v0.9.10 simulator using a vehicle with sensors identical to the KITTI dataset. The vehicle thus has a Velodyne HDL64 LiDAR positioned in the middle of the roof and two color cameras similar to Point Grey Flea 2. The positions of the LiDAR and cameras are the same as the setup used in KITTI. The objective of this dataset is to test approaches of semantic segmentation LiDAR and/or images, odometry LiDAR and/or image in synthetic data and to compare with the results obtained on real data like KITTI. This dataset thus makes it possible to improve transfer learning methods from a synthetic dataset to a real dataset. We created 7 sequences with 5000 frames in each sequence in the 7 maps of CARLA providing different environments (city, suburban area, mountain, rural area, highway...). The dataset is available at: http://npm3d.fr