Open Access

The current state-of-the-art in monocular 3D human pose estimation is heavily influenced by weakly supervised methods. These allow 2D labels to be used to learn effective 3D human pose recovery either directly from images or via 2D-to-3D pose uplifting. In this paper we present a detailed analysis of the most commonly used simplified projection models, which relate the estimated 3D pose representation to 2D labels: normalized perspective and weak perspective projections. Specifically, we derive theoretical lower bound errors for those projection models under the commonly used mean per-joint position error (MPJPE). Additionally, we show how the normalized perspective projection can be replaced to avoid this guaranteed minimal error. We evaluate the derived lower bounds on the most commonly used 3D human pose estimation benchmark datasets. Our results show that both projection models lead to an inherent minimal error between 19.3mm and 54.7mm, even after alignment in position and scale. This is a considerable share when comparing with recent state-of-the-art results. Our paper thus establishes a theoretical baseline that shows the importance of suitable projection models in weakly supervised 3D human pose estimation.

Image manipulation is of historic importance. Ever since the advent of photography, pictures have been manipulated for various reasons. Historic rulers often used image manipulation techniques for the purpose of self-portrayal or propaganda. In many cases, the goal is to manipulate human behaviour by spreading credible misinformation. Photographs, by their nature, portray the real world and as such are more credible to humans. However, image manipulation may not only serve evil purposes. In this thesis, we propose and analyse methods for image manipulation that serve a positive purpose. Specifically, we treat image manipulation as a tool for solving other tasks. For this, we model image manipulation as an image-to-image translation (I2I) task, i.e., a system that receives an image as input and outputs a manipulated version of the input. We propose multiple I2I based methods. We demonstrate that I2I based image manipulation methods can be used to reduce motion blur in videos. Second, we show that I2I based image manipulation methods can be used for domain adaptation and domain extension. Specifically, we present a method that significantly improves the learning of semantic segmentation from synthetic source data. The same technique can be applied to learning nighttime semantic segmentation from daylight images. Next, we show that I2I can be used to enable weakly supervised object segmentation. We show that each individual task requires and allows for different levels of supervision during the training of deep models in order to achieve best performance. We discuss the importance of maintaining control over the output of such methods and show that, with reduced levels of supervision, methods for maintaining stability during training and for establishing control over the output of a system become increasingly important. We propose multiple methods that solve the issues that arise in such systems. Finally, we demonstrate that our proposed mechanisms for control can be adapted to synthesise images from scratch.

Unsupervised domain adaptation is a promising technique for computer vision tasks, especially when annotating large amounts of data is very costly and time-consuming, as in semantic segmentation. Here it is attractive to train neural networks on simulated data and fit them to real data on which the models are to be used. In this paper, we propose a consistency regularization method for domain adaptation in semantic segmentation that combines pseudo-labels and strong per- turbations. We analyse the impact of two simple perturbations, dropout and image mixing, and show how they contribute enormously to the final performance. Experiments and ablation studies demonstrate that our simple approach achieves strong results on relevant synthetic-to-real domain adaptation benchmarks.