Rig-Robust BEV Perception via 3D Gaussian Splatting and Cyclic Self-Supervision

Camera rigs for Advanced Driver Assistance Systems (ADAS) are subject to design choices of each vehicle variant; therefore, ADAS algorithms should be rig-invariant or tunable with limited additional data. Recently, the ADAS community adopted BEV (Bird’s Eye View) perception and planning systems due to their scalability to heterogeneous sensors, computational efficiency, and ego-centric perception, but BEV methods remain highly rig-dependent. To verify this, we study a camera-only BEV model, GaussianLSS, which lifts images into a 3D Gaussian scene and predicts BEV maps. We observe a sharp performance drop under a Sedan-to-SUV rig change, with a similar degradation in the reverse direction. Overlap between predicted and ground-truth BEV maps, measured by intersection-over-union (IoU), falls from 38.42% to 18.30% when trained on Sedan and tested on SUV. To address this, we propose a training-time extension to GaussianLSS that improves cross-rig robustness without using any target-rig images or BEV labels. Starting from sedan images, GaussianLSS lifts features into a 3D Gaussian scene, which we render into rig-aware perspective features for both Sedan and SUV. A lightweight decoder is trained only on sedan perspective features to reconstruct sedan RGB (re-projecting features back to the input image) and is then frozen. During cyclic fine-tuning, SUV perspective features rendered from the same Gaussians are passed through the frozen decoder to produce pseudo-SUV views that provide self-supervision for the BEV model in world/BEV space. This increases SUV performance to 23.46% IoU while keeping Sedan performance near its upper bound, and to 25.15% when the decoder additionally sees 20% SUV RGB (the BEV model still uses no SUV labels). The decoder is not used at inference, keeping deployment cost unchanged while suggests a path toward reducing the need for per-vehicle data collection and annotation.