Robust and Explainable 3D Computer Vision

计算机视觉越来越依赖深度学习，而深度学习是脆弱的、不透明的，而且在没有任何警告的情况下会灾难性地失败。该项目旨在通过开发3D几何和图像数据的图形表示的新理论、分层图形简化和专门为深入学习几何图形而设计的新模块来解决这些问题。利用这些模块，它的目标是设计用于自我监督学习的图卷积网络结构，该结构对故障具有健壮性，并为目标检测和场景分割提供可解释的决策。预计这些结果将推动稳健深度学习的理论，并使3D地图、测量、基础设施监测、交通和机器人行业受益。

Computer vision is increasingly relying on deep learning which is fragile, opaque and fails catastrophically without warning. This project aims to address these problems by developing new theory in graph representation of 3D geometric and image data, hierarchical graph simplification and novel modules designed specifically for deep learning over geometric graphs. Using these modules, it aims to design graph convolutional network architectures for self-supervised learning that are robust to failures and provide explainable decisions for object detection and scene segmentation. The outcomes are expected to advance theory in robust deep learning and benefit 3D mapping, surveying, infrastructure monitoring, transport and robotics industries.