UC Riverside

Scene understanding is an important aspect of computer vision, encompassing a variety of tasks such as image classification, object detection, scene graph generation, and action recognition. With the advances of deep learning models, scene understanding has seen significant improvements in accuracy, robustness, and efficiency, enabling more sophisticated and reliable applications in automated sports analytics, visual relationship detection, and dynamic neural networks. Deep learning techniques that perform on standardized benchmarks often struggle when applied in real-world scenarios where the environment variables are often unconstrained and diverse, resulting in challenges that can hinder their accuracy and generalizability. This thesis delves into the evolution of scene understanding through deep learning, and presents novel approaches for specialized tasks like jersey number detection, scene graph generation, and dynamically throttleable neural networks. The developed techniques enhance the deep model’s ability to learn robust and transferable representations, enabling better generalization across diverse visual domains. Both theory and experimentation will be presented.