Vision perception models have become integral to various downstream tasks, including object detection, segmentation, and tracking, with critical applications in security-sensitive domains like autonomous driving. However, research has demonstrated that even advanced vision perception models powered by deep learning remain vulnerable to both common and adversarial perturbations. In the CDF project, Dr Guo introduce a novel approach using continuous representation to achieve robust vision perception.

The method leverages few-shot observations of a scene to render new observations with arbitrary spatial or temporal coordinates. This continuous representation enables us to resample new observations, providing subsequent perception models with enhanced opportunities for reliable predictions. We have developed several groundbreaking works in this direction: