Large scale analysis of dataset and simulation biases in SLAM research

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Current visual Simultaneous Localization and Mapping (SLAM) and localization methods rely on datasets prone to human biases.
• Photo-realistic simulators offer bias-free data but require precise camera calibration, which is often unavailable.

Purpose of the Study:

• To expose human biases in existing SLAM datasets.
• To introduce a user-friendly method for calibrating simulator first-view cameras for visual navigation.
• To analyze and reduce the simulation-to-reality gap in virtual platforms.

Main Methods:

• Identification and analysis of capture bias and negative world bias in benchmark SLAM datasets.
• Development of a novel intrinsic and extrinsic calibration method for simulator cameras.
• Demonstration of the calibration method on MINOS simulator and GTA-V game.
• Analysis of the simulation-to-reality gap and proposal of a gap-reduction technique.

Main Results:

• Two significant biases (capture and negative world bias) were identified in popular SLAM datasets.
• A novel, user-friendly calibration method for simulator cameras was successfully demonstrated.
• The simulation-to-reality gap was analyzed, and a method for its reduction was proposed.
• Visual navigation algorithms showed performance degradation on novel virtual world scenarios.

Conclusions:

• Simulator calibration is crucial for reliable benchmarking of visual navigation.
• Addressing biases and the simulation-to-reality gap is essential for advancing visual SLAM and localization research.
• The proposed method facilitates the use of simulators for unbiased data generation and algorithm validation.