Title: PairedGTA: A Framework for Generating Driving Datasets to Enable Controlled Photometric Shift Analysis

Abstract:

Ensuring the reliability of visual perception systems in autonomous driving across varied environmental scenarios necessitates rigorous performance evaluation. An ideal assessment of these systems under diverse adverse conditions would rely on perfectly paired imagery capturing the identical scene under differing weather or lighting states. Such pairing enables the isolation of photometric shift effects from geometric and semantic variations. However, obtaining such data from real-world sources is notoriously difficult; because camera positions, traffic patterns, and the locations of dynamic entities (such as pedestrians and vehicles) typically change over time, existing datasets usually offer only coarse pairings.

To overcome this limitation, this study presents a data generation framework utilizing a high-fidelity game engine to extract perfectly aligned image pairs. By interfacing with the Grand Theft Auto (GTA) engine via software APIs, the framework alters illumination and meteorological conditions while maintaining the integrity of scene geometry, camera posture, and the specific identity and positioning of dynamic objects. The system procedurally spawns dynamic entities at sampled locations and renders pixel-aligned images across a spectrum of adverse conditions. The utility of this approach is validated through a systematic evaluation of semantic segmentation models, demonstrating that performance degradation in these scenarios can be more accurately linked to photometric shifts rather than uncontrolled changes in semantics or geometry.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC