MSR Thesis Defense

Abstract: Supervised learning for high-level vision tasks has advanced significantly over the last decade. One of the primary driving forces for these improvements has been the availability of vast amounts of labeled data. However, annotating data is an expensive and time-consuming process. For example, densely segmenting a natural scene image takes approximately 30 minutes. This mode [...]

Abstract: We make an important connection to existing results in econometrics to describe an alternative formulation of inverse reinforcement learning (IRL). In particular, we describe an algorithm to solve the IRL problem, using easy-to-compute estimates of the Conditional Choice Probability (CCP) vector, which is the policy function of an expert integrated over factors econometricians cannot [...]

In this thesis, we explore methods of building dense depth map from monocular video. First, we introduce our multi-view stereo pipeline, which utilizes photometric bundle adjustment for getting accurate depth of textured regions from small motion video. Second, we improve the depth estimation of low-texture region by fusing deep convolutional network predictions. We categorize the [...]

The ability to predict depth from a single image - using recent advances in CNNs - is of increasing interest to the vision community. Unsupervised strategies to learning are particularly appealing as they can utilize much larger and varied monocular video datasets during learning without the need for ground truth depth or stereo. In previous works, separate pose and [...]

This thesis explores learning-based methods in generating human-like lane following and changing behaviors in on-road autonomous driving. We summarize our main contributions as: 1) derive an efficient vision-based end-to-end learning system for on-road driving; 2) propose a novel attention-based learning architecture with sub-action space to obtain lane changing behavior using a deep reinforcement learning algorithm; [...]

Abstract: In the spectrum of vision-based autonomous driving, vanilla end-to-end models are not interpretable and suboptimal in performance, while mediated perception models require additional intermediate representations such as segmentation masks or detection bounding boxes, whose annotation can be prohibitively expensive as we move to a larger scale. More critically, all prior works fail to deal with the notorious [...]

Despite significant research in the area, reconstruction of multiple dynamic rigid objects (eg. vehicles) observed from wide-baseline, uncalibrated and unsynchronized cameras, remains hard. On one hand, feature tracking works well within each view but is hard to correspond across multiple cameras with limited overlap in fields of view or due to occlusions. On the other [...]

In this thesis, we explore sampling-based trajectory optimization applied to search for objects of interest in constrained environments (e.g., a UAV searching for a target in the presence of obstacles). We consider two search scenarios: in the first scenario, accurate prior information distribution of the possible locations of the objects of interest is available, thus [...]

Abstract: Recent success in machine vision has been largely driven by advanced computer vision methods, most commonly known as deep learning based methods. While we have seen tremendous performance improvements in machine visual tasks, such as object categorization and segmentation, there remain two major issues in deep learning. Firstly, deep networks have been largely unable [...]