Matching Items (11)
Filtering by
- All Subjects: Computer vision
- Creators: Jayasuriya, Suren
Description
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.
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.
ContributorsGunnam, Sridhar (Author) / LiKamWa, Robert (Thesis advisor) / Turaga, Pavan (Committee member) / Jayasuriya, Suren (Committee member) / Arizona State University (Publisher)
Created2018
Description
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.
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.
ContributorsDodge, Samuel Fuller (Author) / Karam, Lina (Thesis advisor) / Jayasuriya, Suren (Committee member) / Li, Baoxin (Committee member) / Turaga, Pavan (Committee member) / Arizona State University (Publisher)
Created2018
Description
Visual navigation is a useful and important task for a variety of applications. As the prevalence of robots increase, there is an increasing need for energy-efficient navigation methods as well. Many aspects of efficient visual navigation algorithms have been implemented in the literature, but there is a lack of work on evaluation of the efficiency of the image sensors. In this thesis, two methods are evaluated: adaptive image sensor quantization for traditional camera pipelines as well as new event-based sensors for low-power computer vision.The first contribution in this thesis is an evaluation of performing varying levels of sensor linear and logarithmic quantization with the task of visual simultaneous localization and mapping (SLAM). This unconventional method can provide efficiency benefits with a trade off between accuracy of the task and energy-efficiency. A new sensor quantization method, gradient-based quantization, is introduced to improve the accuracy of the task. This method only lowers the bit level of parts of the image that are less likely to be important in the SLAM algorithm since lower bit levels signify better energy-efficiency, but worse task accuracy. The third contribution is an evaluation of the efficiency and accuracy of event-based camera intensity representations for the task of optical flow. The results of performing a learning based optical flow are provided for each of five different reconstruction methods along with ablation studies. Lastly, the challenges of an event feature-based SLAM system are presented with results demonstrating the necessity for high quality and high resolution event data. The work in this thesis provides studies useful for examining tradeoffs for an efficient visual navigation system with traditional and event vision sensors. The results of this thesis also provide multiple directions for future work.
ContributorsChristie, Olivia Catherine (Author) / Jayasuriya, Suren (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Yang, Yezhou (Committee member) / Arizona State University (Publisher)
Created2022
Description
Computer vision is becoming an essential component of embedded system applications such as smartphones, wearables, autonomous systems and internet-of-things (IoT). These applications are generally deployed into environments with limited energy, memory bandwidth and computational resources. This trend is driving the development of energy-effi cient image processing solutions from sensing to computation. In this thesis, diff erent alternatives are explored to implement energy-efficient computer vision systems. First, I present a fi eld programmable gate array (FPGA) implementation of an adaptive subsampling algorithm for region-of-interest (ROI) -based object tracking. By implementing the computationally intensive sections of this algorithm on an FPGA, I aim to offl oad computing resources from energy-ineffi cient graphics processing units (GPUs) and/or general-purpose central processing units (CPUs). I also present a working system executing this algorithm in near real-time latency implemented on a standalone embedded device. Secondly, I present a neural network-based pipeline to improve the performance of event-based cameras in non-ideal optical conditions. Event-based cameras or dynamic vision sensors (DVS) are bio-inspired sensors that measure logarithmic per-pixel brightness changes in a scene. Their advantages include high dynamic range, low latency and ultra-low power when compared to standard frame-based cameras. Several tasks have been proposed to take advantage of these novel sensors but they rely on perfectly calibrated optical lenses that are in-focus. In this work I propose a methodto reconstruct events captured with an out-of-focus event-camera so they can be fed into an intensity reconstruction task. The network is trained with a dataset generated by simulating defocus blur in sequences from object tracking datasets such as LaSOT and OTB100. I also test the generalization performance of this network in scenes captured with a DAVIS event-based sensor equipped with an out-of-focus lens.
ContributorsTorres Muro, Victor Isaac (Author) / Jayasuriya, Suren (Thesis advisor) / Spanias, Andreas (Committee member) / Seo, Jae-Sun (Committee member) / Arizona State University (Publisher)
Created2022
Description
Over the last decade, deep neural networks also known as deep learning, combined with large databases and specialized hardware for computation, have made major strides in important areas such as computer vision, computational imaging and natural language processing. However, such frameworks currently suffer from some drawbacks. For example, it is generally not clear how the architectures are to be designed for different applications, or how the neural networks behave under different input perturbations and it is not easy to make the internal representations and parameters more interpretable. In this dissertation, I propose building constraints into feature maps, parameters and and design of algorithms involving neural networks for applications in low-level vision problems such as compressive imaging and multi-spectral image fusion, and high-level inference problems including activity and face recognition. Depending on the application, such constraints can be used to design architectures which are invariant/robust to certain nuisance factors, more efficient and, in some cases, more interpretable. Through extensive experiments on real-world datasets, I demonstrate these advantages of the proposed methods over conventional frameworks.
ContributorsLohit, Suhas Anand (Author) / Turaga, Pavan (Thesis advisor) / Spanias, Andreas (Committee member) / Li, Baoxin (Committee member) / Jayasuriya, Suren (Committee member) / Arizona State University (Publisher)
Created2019
Description
Recently, a well-designed and well-trained neural network can yield state-of-the-art results across many domains, including data mining, computer vision, and medical image analysis. But progress has been limited for tasks where labels are difficult or impossible to obtain. This reliance on exhaustive labeling is a critical limitation in the rapid deployment of neural networks. Besides, the current research scales poorly to a large number of unseen concepts and is passively spoon-fed with data and supervision.
To overcome the above data scarcity and generalization issues, in my dissertation, I first propose two unsupervised conventional machine learning algorithms, hyperbolic stochastic coding, and multi-resemble multi-target low-rank coding, to solve the incomplete data and missing label problem. I further introduce a deep multi-domain adaptation network to leverage the power of deep learning by transferring the rich knowledge from a large-amount labeled source dataset. I also invent a novel time-sequence dynamically hierarchical network that adaptively simplifies the network to cope with the scarce data.
To learn a large number of unseen concepts, lifelong machine learning enjoys many advantages, including abstracting knowledge from prior learning and using the experience to help future learning, regardless of how much data is currently available. Incorporating this capability and making it versatile, I propose deep multi-task weight consolidation to accumulate knowledge continuously and significantly reduce data requirements in a variety of domains. Inspired by the recent breakthroughs in automatically learning suitable neural network architectures (AutoML), I develop a nonexpansive AutoML framework to train an online model without the abundance of labeled data. This work automatically expands the network to increase model capability when necessary, then compresses the model to maintain the model efficiency.
In my current ongoing work, I propose an alternative method of supervised learning that does not require direct labels. This could utilize various supervision from an image/object as a target value for supervising the target tasks without labels, and it turns out to be surprisingly effective. The proposed method only requires few-shot labeled data to train, and can self-supervised learn the information it needs and generalize to datasets not seen during training.
To overcome the above data scarcity and generalization issues, in my dissertation, I first propose two unsupervised conventional machine learning algorithms, hyperbolic stochastic coding, and multi-resemble multi-target low-rank coding, to solve the incomplete data and missing label problem. I further introduce a deep multi-domain adaptation network to leverage the power of deep learning by transferring the rich knowledge from a large-amount labeled source dataset. I also invent a novel time-sequence dynamically hierarchical network that adaptively simplifies the network to cope with the scarce data.
To learn a large number of unseen concepts, lifelong machine learning enjoys many advantages, including abstracting knowledge from prior learning and using the experience to help future learning, regardless of how much data is currently available. Incorporating this capability and making it versatile, I propose deep multi-task weight consolidation to accumulate knowledge continuously and significantly reduce data requirements in a variety of domains. Inspired by the recent breakthroughs in automatically learning suitable neural network architectures (AutoML), I develop a nonexpansive AutoML framework to train an online model without the abundance of labeled data. This work automatically expands the network to increase model capability when necessary, then compresses the model to maintain the model efficiency.
In my current ongoing work, I propose an alternative method of supervised learning that does not require direct labels. This could utilize various supervision from an image/object as a target value for supervising the target tasks without labels, and it turns out to be surprisingly effective. The proposed method only requires few-shot labeled data to train, and can self-supervised learn the information it needs and generalize to datasets not seen during training.
ContributorsZhang, Jie (Author) / Wang, Yalin (Thesis advisor) / Liu, Huan (Committee member) / Stonnington, Cynthia (Committee member) / Liang, Jianming (Committee member) / Yang, Yezhou (Committee member) / Arizona State University (Publisher)
Created2020
Description
In recent years, the widespread use of deep neural networks (DNNs) has facilitated great improvements in performance for computer vision tasks like image classification and object recognition. In most realistic computer vision applications, an input image undergoes some form of image distortion such as blur and additive noise during image acquisition or transmission. Deep networks trained on pristine images perform poorly when tested on such distortions. DNN predictions have also been shown to be vulnerable to carefully crafted adversarial perturbations. Specifically, so-called universal adversarial perturbations are image-agnostic perturbations that can be added to any image and can fool a target network into making erroneous predictions. This work proposes selective DNN feature regeneration to improve the robustness of existing DNNs to image distortions and universal adversarial perturbations.
In the context of common naturally occurring image distortions, a metric is proposed to identify the most susceptible DNN convolutional filters and rank them in order of the highest gain in classification accuracy upon correction. The proposed approach called DeepCorrect applies small stacks of convolutional layers with residual connections at the output of these ranked filters and trains them to correct the most distortion-affected filter activations, whilst leaving the rest of the pre-trained filter outputs in the network unchanged. Performance results show that applying DeepCorrect models for common vision tasks significantly improves the robustness of DNNs against distorted images and outperforms other alternative approaches.
In the context of universal adversarial perturbations, departing from existing defense strategies that work mostly in the image domain, a novel and effective defense which only operates in the DNN feature domain is presented. This approach identifies pre-trained convolutional features that are most vulnerable to adversarial perturbations and deploys trainable feature regeneration units which transform these DNN filter activations into resilient features that are robust to universal perturbations. Regenerating only the top 50% adversarially susceptible activations in at most 6 DNN layers and leaving all remaining DNN activations unchanged can outperform existing defense strategies across different network architectures and across various universal attacks.
In the context of common naturally occurring image distortions, a metric is proposed to identify the most susceptible DNN convolutional filters and rank them in order of the highest gain in classification accuracy upon correction. The proposed approach called DeepCorrect applies small stacks of convolutional layers with residual connections at the output of these ranked filters and trains them to correct the most distortion-affected filter activations, whilst leaving the rest of the pre-trained filter outputs in the network unchanged. Performance results show that applying DeepCorrect models for common vision tasks significantly improves the robustness of DNNs against distorted images and outperforms other alternative approaches.
In the context of universal adversarial perturbations, departing from existing defense strategies that work mostly in the image domain, a novel and effective defense which only operates in the DNN feature domain is presented. This approach identifies pre-trained convolutional features that are most vulnerable to adversarial perturbations and deploys trainable feature regeneration units which transform these DNN filter activations into resilient features that are robust to universal perturbations. Regenerating only the top 50% adversarially susceptible activations in at most 6 DNN layers and leaving all remaining DNN activations unchanged can outperform existing defense strategies across different network architectures and across various universal attacks.
ContributorsBorkar, Tejas Shyam (Author) / Karam, Lina J (Thesis advisor) / Turaga, Pavan (Committee member) / Jayasuriya, Suren (Committee member) / Tepedelenlioğlu, Cihan (Committee member) / Arizona State University (Publisher)
Created2020
Description
Atmospheric turbulence distorts the path of light passing through the air. When capturing images at long range, the effects of this turbulence can cause substantial geometric distortion and blur in images and videos, degrading image quality. These become more pronounced with greater turbulence, scaling with the refractive index structure constant, Cn2. Removing effects of atmospheric turbulence in images has a range of applications from astronomical imaging to surveillance. Thus, there is great utility in transforming a turbulent image into a ``clean image" undegraded by turbulence. However, as the turbulence is space- and time-variant and statistically random, no closed-form solution exists for a function that performs this transformation. Prior attempts to approximate the solution include spatio-temporal models and lucky frames models, which require many images to provide a good approximation, and supervised neural networks, which rely on large amounts of simulated or difficult-to-acquire real training data and can struggle to generalize. The first contribution in this thesis is an unsupervised neural-network-based model to perform image restoration for atmospheric turbulence with state-of-the-art performance. The model consists of a grid deformer, which produces an estimated distortion field, and an image generator, which estimates the distortion-free image. This model is transferable across different datasets; its efficacy is demonstrated across multiple datasets and on both air and water turbulence. The second contribution is a supervised neural network to predict Cn2 directly from the warp field. This network was trained on a wide range of Cn2 values and estimates Cn2 with relatively good accuracy. When used on the warp field produced by the unsupervised model, this allows for a Cn2 estimate requiring only a few images without any prior knowledge of ground truth or information about the turbulence.
ContributorsWhyte, Cameron (Author) / Jayasuriya, Suren (Thesis advisor) / Espanol, Malena (Thesis advisor) / Speyer, Gil (Committee member) / Arizona State University (Publisher)
Created2021
Description
Machine learning (ML) and deep learning (DL) has become an intrinsic part of multiple fields. The ability to solve complex problems makes machine learning a panacea. In the last few years, there has been an explosion of data generation, which has greatly improvised machine learning models. But this comes with a cost of high computation, which invariably increases power usage and cost of the hardware. In this thesis we explore applications of ML techniques, applied to two completely different fields - arts, media and theater and urban climate research using low-cost and low-powered edge devices. The multi-modal chatbot uses different machine learning techniques: natural language processing (NLP) and computer vision (CV) to understand inputs of the user and accordingly perform in the play and interact with the audience. This system is also equipped with other interactive hardware setups like movable LED systems, together they provide an experiential theatrical play tailored to each user. I will discuss how I used edge devices to achieve this AI system which has created a new genre in theatrical play. I will then discuss MaRTiny, which is an AI-based bio-meteorological system that calculates mean radiant temperature (MRT), which is an important parameter for urban climate research. It is also equipped with a vision system that performs different machine learning tasks like pedestrian and shade detection. The entire system costs around $200 which can potentially replace the existing setup worth $20,000. I will further discuss how I overcame the inaccuracies in MRT value caused by the system, using machine learning methods. These projects although belonging to two very different fields, are implemented using edge devices and use similar ML techniques. In this thesis I will detail out different techniques that are shared between these two projects and how they can be used in several other applications using edge devices.
ContributorsKulkarni, Karthik Kashinath (Author) / Jayasuriya, Suren (Thesis advisor) / Middel, Ariane (Thesis advisor) / Yu, Hongbin (Committee member) / Arizona State University (Publisher)
Created2021
Description
Commonly, image processing is handled on a CPU that is connected to the image sensor by a wire. In these far-sensor processing architectures, there is energy loss associated with sending data across an interconnect from the sensor to the CPU. In an effort to increase energy efficiency, near-sensor processing architectures have been developed, in which the sensor and processor are stacked directly on top of each other. This reduces energy loss associated with sending data off-sensor. However, processing near the image sensor causes the sensor to heat up. Reports of thermal noise in near-sensor processing architectures motivated us to study how temperature affects image quality on a commercial image sensor and how thermal noise affects computer vision task accuracy. We analyzed image noise across nine different temperatures and three sensor configurations to determine how image noise responds to an increase in temperature. Ultimately, our team used this information, along with transient analysis of a stacked image sensor’s thermal behavior, to advise thermal management strategies that leverage the benefits of near-sensor processing and prevent accuracy loss at problematic temperatures.
ContributorsJones, Britton Steele (Author) / LiKamWa, Robert (Thesis director) / Jayasuriya, Suren (Committee member) / Watts College of Public Service & Community Solut (Contributor) / Electrical Engineering Program (Contributor, Contributor) / Barrett, The Honors College (Contributor)
Created2020-12