Learning From Multi-Frame Data

Multi-frame data-driven methods bear the promise that aggregating multiple observations leads to better estimates of target quantities than a single (still) observation. This thesis examines how data-driven approaches such as deep neural networks should be constructed to improve over single-frame-based counterparts. Besides algorithmic changes, as for example in the design of artificial neural network architectures or the algorithm itself, such an examination is inextricably linked with the consideration of the synthesis of synthetic training data in meaningful size (even if no annotations are available) and quality (if real ground-truth acquisition is not possible), which capture all temporal effects with high fidelity. We start with the introduction of a new algorithm to accelerate a nonparametric learning algorithm by using a GPU adapted implementation to search for the nearest neighbor. While the approaches known so far are clearly surpassed, this empirically reveals that the data generated can be managed within a reasonable time and that several inputs can be processed in parallel even under hardware restrictions. Based on a learning-based solution, we introduce a novel training protocol to bridge the need for carefully curated training data and demonstrate better performance and robustness than a non-parametric search for the nearest neighbor via temporal video alignments. Effective learning in the absence of labels is required when dealing with larger amounts of data that are easy to capture but not feasible or at least costly to label. In addition, we show new ways to generate plausible and realistic synthesized data and their inevitability when it comes to closing the gap to expensive and almost infeasible real-world acquisition. These eventually achieve state-of-the-art results in classical image processing tasks such as reflection removal and video deblurring