Disentangling latent spaces is an important research direction in the interpretability of unsupervised machine learning. Several recent works using deep learning are very effective at producing disentangled representations. However, in the unsupervised setting, there is no way to pre-specify which part of the latent space captures specific factors of variations. While this is generally a hard problem because of the non-existence of analytical expressions to capture these variations, there are certain factors like geometric

transforms that can be expressed analytically. Furthermore, in existing frameworks, the disentangled values are also not interpretable. The focus of this work is to disentangle these geometric factors of variations (which turn out to be nuisance factors for many applications) from the semantic content of the signal in an interpretable manner which in turn makes the features more discriminative. Experiments are designed to show the modularity of the approach with other disentangling strategies as well as on multiple one-dimensional (1D) and two-dimensional (2D) datasets, clearly indicating the efficacy of the proposed approach.

Despite the fact that machine learning supports the development of computer vision applications by shortening the development cycle, finding a general learning algorithm that solves a wide range of applications is still bounded by the ”no free lunch theorem”. The search for the right algorithm to solve a specific problem is driven by the problem itself, the data availability and many other requirements.

Automated visual inspection (AVI) systems represent a major part of these challenging computer vision applications. They are gaining growing interest in the manufacturing industry to detect defective products and keep these from reaching customers. The process of defect detection and classification in semiconductor units is challenging due to different acceptable variations that the manufacturing process introduces. Other variations are also typically introduced when using optical inspection systems due to changes in lighting conditions and misalignment of the imaged units, which makes the defect detection process more challenging.

In this thesis, a BagStack classification framework is proposed, which makes use of stacking and bagging concepts to handle both variance and bias errors. The classifier is designed to handle the data imbalance and overfitting problems by adaptively transforming the

multi-class classification problem into multiple binary classification problems, applying a bagging approach to train a set of base learners for each specific problem, adaptively specifying the number of base learners assigned to each problem, adaptively specifying the number of samples to use from each class, applying a novel data-imbalance aware cross-validation technique to generate the meta-data while taking into account the data imbalance problem at the meta-data level and, finally, using a multi-response random forest regression classifier as a meta-classifier. The BagStack classifier makes use of multiple features to solve the defect classification problem. In order to detect defects, a locally adaptive statistical background modeling is proposed. The proposed BagStack classifier outperforms state-of-the-art image classification techniques on our dataset in terms of overall classification accuracy and average per-class classification accuracy. The proposed detection method achieves high performance on the considered dataset in terms of recall and precision.

Non-line-of-sight (NLOS) imaging of objects not visible to either the camera or illumina-

tion source is a challenging task with vital applications including surveillance and robotics.

Recent NLOS reconstruction advances have been achieved using time-resolved measure-

ments. Acquiring these time-resolved measurements requires expensive and specialized

detectors and laser sources. In work proposes a data-driven approach for NLOS 3D local-

ization requiring only a conventional camera and projector. The localisation is performed

using a voxelisation and a regression problem. Accuracy of greater than 90% is achieved

in localizing a NLOS object to a 5cm × 5cm × 5cm volume in real data. By adopting

the regression approach an object of width 10cm to localised to approximately 1.5cm. To

generalize to line-of-sight (LOS) scenes with non-planar surfaces, an adaptive lighting al-

gorithm is adopted. This algorithm, based on radiosity, identifies and illuminates scene

patches in the LOS which most contribute to the NLOS light paths, and can factor in sys-

tem power constraints. Improvements ranging from 6%-15% in accuracy with a non-planar

LOS wall using adaptive lighting is reported, demonstrating the advantage of combining

the physics of light transport with active illumination for data-driven NLOS imaging.

Speech is generated by articulators acting on

a phonatory source. Identification of this

phonatory source and articulatory geometry are

individually challenging and ill-posed

problems, called speech separation and

articulatory inversion, respectively.

There exists a trade-off

between decomposition and recovered

articulatory geometry due to multiple

possible mappings between an

articulatory configuration

and the speech produced. However, if measurements

are obtained only from a microphone sensor,

they lack any invasive insight and add

additional challenge to an already difficult

problem.

A joint non-invasive estimation

strategy that couples articulatory and

phonatory knowledge would lead to better

articulatory speech synthesis. In this thesis,

a joint estimation strategy for speech

separation and articulatory geometry recovery

is studied. Unlike previous

periodic/aperiodic decomposition methods that

use stationary speech models within a

frame, the proposed model presents a

non-stationary speech decomposition method.

A parametric glottal source model and an

articulatory vocal tract response are

represented in a dynamic state space formulation.

The unknown parameters of the

speech generation components are estimated

using sequential Monte Carlo methods

under some specific assumptions.

The proposed approach is compared with other

glottal inverse filtering methods,

including iterative adaptive inverse filtering,

state-space inverse filtering, and

the quasi-closed phase method.

Motion estimation is a core task in computer vision and many applications utilize optical flow methods as fundamental tools to analyze motion in images and videos. Optical flow is the apparent motion of objects in image sequences that results from relative motion between the objects and the imaging perspective. Today, optical flow fields are utilized to solve problems in various areas such as object detection and tracking, interpolation, visual odometry, etc. In this dissertation, three problems from different areas of computer vision and the solutions that make use of modified optical flow methods are explained.

The contributions of this dissertation are approaches and frameworks that introduce i) a new optical flow-based interpolation method to achieve minimally divergent velocimetry data, ii) a framework that improves the accuracy of change detection algorithms in synthetic aperture radar (SAR) images, and iii) a set of new methods to integrate Proton Magnetic Resonance Spectroscopy (1HMRSI) data into threedimensional (3D) neuronavigation systems for tumor biopsies.

In the first application an optical flow-based approach for the interpolation of minimally divergent velocimetry data is proposed. The velocimetry data of incompressible fluids contain signals that describe the flow velocity. The approach uses the additional flow velocity information to guide the interpolation process towards reduced divergence in the interpolated data.

In the second application a framework that mainly consists of optical flow methods and other image processing and computer vision techniques to improve object extraction from synthetic aperture radar images is proposed. The proposed framework is used for distinguishing between actual motion and detected motion due to misregistration in SAR image sets and it can lead to more accurate and meaningful change detection and improve object extraction from a SAR datasets.

In the third application a set of new methods that aim to improve upon the current state-of-the-art in neuronavigation through the use of detailed three-dimensional (3D) 1H-MRSI data are proposed. The result is a progressive form of online MRSI-guided neuronavigation that is demonstrated through phantom validation and clinical application.

Human movement is a complex process influenced by physiological and psychological factors. The execution of movement is varied from person to person, and the number of possible strategies for completing a specific movement task is almost infinite. Different choices of strategies can be perceived by humans as having different degrees of quality, and the quality can be defined with regard to aesthetic, athletic, or health-related ratings. It is useful to measure and track the quality of a person's movements, for various applications, especially with the prevalence of low-cost and portable cameras and sensors today. Furthermore, based on such measurements, feedback systems can be designed for people to practice their movements towards certain goals. In this dissertation, I introduce symmetry as a family of measures for movement quality, and utilize recent advances in computer vision and differential geometry to model and analyze different types of symmetry in human movements. Movements are modeled as trajectories on different types of manifolds, according to the representations of movements from sensor data. The benefit of such a universal framework is that it can accommodate different existing and future features that describe human movements. The theory and tools developed in this dissertation will also be useful in other scientific areas to analyze symmetry from high-dimensional signals.

Mixture of experts is a machine learning ensemble approach that consists of individual models that are trained to be ``experts'' on subsets of the data, and a gating network that provides weights to output a combination of the expert predictions. Mixture of experts models do not currently see wide use due to difficulty in training diverse experts and high computational requirements. This work presents modifications of the mixture of experts formulation that use domain knowledge to improve training, and incorporate parameter sharing among experts to reduce computational requirements.

First, this work presents an application of mixture of experts models for quality robust visual recognition. First it is shown that human subjects outperform deep neural networks on classification of distorted images, and then propose a model, MixQualNet, that is more robust to distortions. The proposed model consists of ``experts'' that are trained on a particular type of image distortion. The final output of the model is a weighted sum of the expert models, where the weights are determined by a separate gating network. The proposed model also incorporates weight sharing to reduce the number of parameters, as well as increase performance.



Second, an application of mixture of experts to predict visual saliency is presented. A computational saliency model attempts to predict where humans will look in an image. In the proposed model, each expert network is trained to predict saliency for a set of closely related images. The final saliency map is computed as a weighted mixture of the expert networks' outputs, with weights determined by a separate gating network. The proposed model achieves better performance than several other visual saliency models and a baseline non-mixture model.

Finally, this work introduces a saliency model that is a weighted mixture of models trained for different levels of saliency. Levels of saliency include high saliency, which corresponds to regions where almost all subjects look, and low saliency, which corresponds to regions where some, but not all subjects look. The weighted mixture shows improved performance compared with baseline models because of the diversity of the individual model predictions.

Generating real-world content for VR is challenging in terms of capturing and processing at high resolution and high frame-rates. The content needs to represent a truly immersive experience, where the user can look around in 360-degree view and perceive the depth of the scene. The existing solutions only capture and offload the compute load to the server. But offloading large amounts of raw camera feeds takes longer latencies and poses difficulties for real-time applications. By capturing and computing on the edge, we can closely integrate the systems and optimize for low latency. However, moving the traditional stitching algorithms to battery constrained device needs at least three orders of magnitude reduction in power. We believe that close integration of capture and compute stages will lead to reduced overall system power.

We approach the problem by building a hardware prototype and characterize the end-to-end system bottlenecks of power and performance. The prototype has 6 IMX274 cameras and uses Nvidia Jetson TX2 development board for capture and computation. We found that capturing is bottlenecked by sensor power and data-rates across interfaces, whereas compute is limited by the total number of computations per frame. Our characterization shows that redundant capture and redundant computations lead to high power, huge memory footprint, and high latency. The existing systems lack hardware-software co-design aspects, leading to excessive data transfers across the interfaces and expensive computations within the individual subsystems. Finally, we propose mechanisms to optimize the system for low power and low latency. We emphasize the importance of co-design of different subsystems to reduce and reuse the data. For example, reusing the motion vectors of the ISP stage reduces the memory footprint of the stereo correspondence stage. Our estimates show that pipelining and parallelization on custom FPGA can achieve real time stitching.

The performance of most of the visual computing tasks depends on the quality of the features extracted from the raw data. Insightful feature representation increases the performance of many learning algorithms by exposing the underlying explanatory factors of the output for the unobserved input. A good representation should also handle anomalies in the data such as missing samples and noisy input caused by the undesired, external factors of variation. It should also reduce the data redundancy. Over the years, many feature extraction processes have been invented to produce good representations of raw images and videos.

The feature extraction processes can be categorized into three groups. The first group contains processes that are hand-crafted for a specific task. Hand-engineering features requires the knowledge of domain experts and manual labor. However, the feature extraction process is interpretable and explainable. Next group contains the latent-feature extraction processes. While the original feature lies in a high-dimensional space, the relevant factors for a task often lie on a lower dimensional manifold. The latent-feature extraction employs hidden variables to expose the underlying data properties that cannot be directly measured from the input. Latent features seek a specific structure such as sparsity or low-rank into the derived representation through sophisticated optimization techniques. The last category is that of deep features. These are obtained by passing raw input data with minimal pre-processing through a deep network. Its parameters are computed by iteratively minimizing a task-based loss.

In this dissertation, I present four pieces of work where I create and learn suitable data representations. The first task employs hand-crafted features to perform clinically-relevant retrieval of diabetic retinopathy images. The second task uses latent features to perform content-adaptive image enhancement. The third task ranks a pair of images based on their aestheticism. The goal of the last task is to capture localized image artifacts in small datasets with patch-level labels. For both these tasks, I propose novel deep architectures and show significant improvement over the previous state-of-art approaches. A suitable combination of feature representations augmented with an appropriate learning approach can increase performance for most visual computing tasks.

Computer Vision as a eld has gone through signicant changes in the last decade.

The eld has seen tremendous success in designing learning systems with hand-crafted

features and in using representation learning to extract better features. In this dissertation

some novel approaches to representation learning and task learning are studied.

Multiple-instance learning which is generalization of supervised learning, is one

example of task learning that is discussed. In particular, a novel non-parametric k-

NN-based multiple-instance learning is proposed, which is shown to outperform other

existing approaches. This solution is applied to a diabetic retinopathy pathology

detection problem eectively.

In cases of representation learning, generality of neural features are investigated

rst. This investigation leads to some critical understanding and results in feature

generality among datasets. The possibility of learning from a mentor network instead

of from labels is then investigated. Distillation of dark knowledge is used to eciently

mentor a small network from a pre-trained large mentor network. These studies help

in understanding representation learning with smaller and compressed networks.