3DLane-TTA: test-time adaptation for monocular 3D lane detection via meta-auxiliary prompt learning

Reliable monocular 3D lane detection is essential for autonomous driving, yet deploying such measurement models across diverse environments remains challenging. Most learning-based methods suffer notable performance degradation when confronted with unseen domains that have mismatched visual conditions. In this paper, we propose 3DLane-TTA, a parameter-efficient test-time adaptation (TTA) framework for monocular 3D lane detection. 3DLane-TTA performs instance-specific refinement through visual prompt tuning, updating fewer than 1% of the model parameters while keeping the backbone frozen. To ensure stable and effective adaptation, we introduce a meta-auxiliary learning strategy that aligns auxiliary gradients with the primary 3D lane detection objective. The framework is further driven by three complementary self-supervised auxiliary tasks, including bird’s-eye-view geometric consistency, feature self-distillation, and masked reconstruction, which enable label-free test-time adaptation using only the current test image. Extensive evaluations on ApolloSim, ONCE-3DLane, and OpenLane demonstrate that 3DLane-TTA generalizes effectively to unseen test data and consistently achieves state-of-the-art performance across diverse and challenging real-world driving scenarios.