Robust Training Framework via Multi-Stage Feature Rectification

Learning robust representations in vision models is essential for reliable performance under diverse real-world conditions, such as weather-induced noise or distribution shifts. In this paper, we propose a novel robust training framework that encourages feature-level rectification directly within the backbone network during training. Unlike existing approaches that rely on external modules or post-processing, our method introduces minimal overhead while enhancing the inherent robustness of the encoder itself. To achieve this, we construct a paired dataset of clean and task-specific noisy images, and apply three complementary training strategies: (1) a reconstruction decoder to align the pixel-space outputs of clean and noisy inputs; (2) contrastive learning to enforce latent similarity between the two views; and (3) a quantization module that constrains latent features to discrete clean representations using vector quantization with a rotation trick. We validate our framework on the KITTI-360 dataset under various weather perturbations, showing significant performance gains in object detection without degrading clean performance. Our approach is lightweight, modular, and applicable to any multi-scale feature-extracting backbone, making it ideal for safety-critical applications such as autonomous driving.