Toward Reproducible Version-Controlled Perception Platforms: Embracing Simplicity in Autonomous Vehicle Dataset Acquisition

Building datasets for autonomous vehicles has become an essential element of robotics research. Numerous datasets were published, pushing the state-of-the-art forward. This article investigates the problem of building reliable perception platforms that can accurately capture and process sensor data, ensuring the integrity and quality of the datasets generated. We propose using version control systems to enhance dataset acquisition's efficiency, reliability, and scalability. We present a method that can launch the system and record data while logging the exact state of the system simultaneously, making the setup and the data generated with it more reliable and reproducible. The main contribution of this paper is a systematic method that applies to existing and operating perception platforms used to collect data. Our framework is based solely on standard tools and is independent of the chosen sensor suite or host system. Implementing our method for existing perception platforms is possible, and to facilitate this, we open-source all the software used to operate our perception platform at: https://www.github.com/ipb-car.