Download or read book Combining Geometry and Learning for Scene Understanding written by Arun Kumar Chockalingam Santha Kumar and published by . This book was released on 2018 with total page 344 pages. Available in PDF, EPUB and Kindle. Book excerpt: When an image is captured, the 3D Euclidean space describing its world is projected onto a 2D plane, effectively losing most pertinent underlying 3D Euclidean geometry information. Consequently, the ultimate goal of any 3D scene understanding system is to recover the lost 3D geometry while deconstructing the semantics of the scene, to perform perceptual decision making tasks such as object detection, pose estimation, shape recovery, etc. Furthermore, the ill-posed nature of the 3D scene recovery problem, where multiple shapes can generate the same image, adds further complexity to an already challenging problem. Significant research has been devoted toward solving the 3D scene recovery problem over the past few decades, with approaches ranging from triangulation and space carving using multiple views of the scene, to using learning-based models to learn semantic priors, to reason and reconstruct the scene. An alternative view of 3D scene understanding is that, given large amounts of data, it is possible to design machines that can automatically learn relevant relationships to perform various vision tasks such as reconstruction, pose prediction, etc. with minimal human supervision and without resorting to complex, manually-designed objective functions. The recent upsurge in deep learning techniques and abundance of data accompanied by the availability of annotations, has resulted in several state-of-the-art learning-based 3D reconstruction models that regress the underlying information in a purely data-driven manner. However, the success of deep learning has come at a hefty price, from the cost of gathering training data to the cost of painstaking labor involved in manual annotation of these data. In light of the above, the goal of this dissertation is to explore both learning-based and geometry-based approaches to 3D scene reconstruction, more specifically, equip learning-based models with geometric reasoning to enable joint scene understanding. This dissertation aims to move away from annotation-intensive learning-based techniques to develop 3D scene reconstruction models that harness the power of geometry and learn from arbitrary data instead of from manually curated 3D datasets, exploit class priors, and most importantly, address the learning and geometric reasoning tasks holistically, to more effectively combat ambiguities in reconstruction and recognition.