Towards Object-Centric Scene Understanding

Visual perception for autonomous agents continues to attract community attention due to the disruptive technologies and the wide applicability of such solutions. Autonomous Driving (AD), a major application in this domain, promises to revolutionize our approach to mobility while bringing critical advantages in limiting accident fatalities. Fueled by recent advances in Deep Learning (DL), more computer vision tasks are being addressed using a learning paradigm. Deep Neural Networks (DNNs) succeeded consistently in pushing performances to unprecedented levels and demonstrating the ability of such approaches to generalize to an increasing number of difficult problems, such as 3D vision tasks. In this thesis, we address two main challenges arising from the current approaches. Namely, the computational complexity of multi-task pipelines, and the increasing need for manual annotations. On the one hand, AD systems need to perceive the surrounding environment on different levels of detail and, subsequently, take timely actions. This multitasking further limits the time available for each perception task. On the other hand, the need for universal generalization of such systems to massively diverse situations requires the use of large-scale datasets covering long-tailed cases. Such requirement renders the use of traditional supervised approaches, despite the data readily available in the AD domain, unsustainable in terms of annotation costs, especially for 3D tasks. Driven by the AD environment nature and the complexity dominated (unlike indoor scenes) by the presence of other scene elements (mainly cars and pedestrians) we focus on the above-mentioned challenges in object-centric tasks. We, then, situate our contributions appropriately in fast-paced literature, while supporting our claims with extensive experimental analysis leveraging up-to-date state-of-the-art results and community-adopted benchmarks.