Toward Efficient and Robust Computer Vision for Large-Scale Edge Applications

The past decade has been witnessing remarkable advancements in computer vision and deep learning algorithms, ushering in a transformative wave of large-scale edge applications across various industries. These image processing methods, however, still encounter numerous challenges when it comes to meeting real-world demands, especially in terms of accuracy and latency at scale. Indeed, striking a balance among efficiency, robustness, and scalability remains a common obstacle. This dissertation investigates these issues in the context of different computer vision tasks, including image classification, semantic segmentation, depth estimation, and object detection. We introduce novel solutions, focusing on utilizing adjustable neural networks, joint multi-task architecture search, and generalized supervision interpolation. The first obstacle revolves around the ability to trade off between speed and accuracy in convolutional neural networks (CNNs) during inference on resource-constrained platforms. Despite their progress, CNNs are typically monolithic at runtime, which can present practical difficulties since computational budgets may vary over time. To address this, we introduce Any-Width Network, an adjustable-width CNN architecture that utilizes a novel Triangular Convolution module to enable fine-grained control over speed and accuracy during inference. The second challenge focuses on the computationally demanding nature of dense prediction tasks such as semantic segmentation and depth estimation. This issue becomes especially problematic for edge platforms with limited resources. To tackle this, we propose a novel and scalable framework named EDNAS. EDNAS leverages the synergistic relationship between Multi-Task Learning and hardware-aware Neural Architecture Search to significantly enhance on-device speed and accuracy of dense predictions. Finally, to improve the robustness of object detection, we introduce a novel data mixing augmentation. While mixing techniques such as Mixup have proven successful in image classification, their application to object detection is non-trivial due to spatial misalignment, foreground/background distinction, and instance multiplicity. To address these issues, we propose a generalized data mixing principle, Supervision Interpolation, and its simple yet effective implementation, LossMix. By addressing these challenges, this dissertation aims to facilitate better efficiency, accuracy, and scalability of computer vision and deep learning algorithms and contribute to the advancement of large-scale edge applications across different domains.Doctor of Philosoph