Rotation-Invariant Self-Supervised Representation Learning for Object Detection in Autonomous Driving

This paper describes a self supervised representation learning approach that can perform robust object detection in out-of-distribution rotated images for autonomous driving task. Keeping in mind the limitations of vehicle sensors we incorporate self-supervised rotation invariance for correcting rotational corruptions. We present a comprehensive analysis of observed label shifts, validate the results using holdout sets and provide explainability for the performance uisng GradCAM appaorach. We demonstrate results on real-world adversarial examples: (Autonomous Scenes - DAWN dataset) that signifi-cantly outperform common methods. The proposed framework outperforms the rotation-affected benchmark by 14.73 points of mAP for the OD task in autonomous driving applications.