Preview

PDF, English - main document Download (39MB) | Lizenz: Learning Instance Segmentation from Sparse Supervision by Wolny, Adrian underlies the terms of Creative Commons Attribution 4.0

Abstract

Instance segmentation is an important task in many domains of automatic image processing, such as self-driving cars, robotics and microscopy data analysis. Recently, deep learning-based algorithms have brought image segmentation close to human performance. However, most existing models rely on dense groundtruth labels for training, which are expensive, time consuming and often require experienced annotators to perform the labeling. Besides the annotation burden, training complex high-capacity neural networks depends upon non-trivial expertise in the choice and tuning of hyperparameters, making the adoption of these models challenging for researchers in other fields.

The aim of this work is twofold. The first is to make the deep learning segmentation methods accessible to non-specialist. The second is to address the dense annotation problem by developing instance segmentation methods trainable with limited groundtruth data. In the first part of this thesis, I bring state-of-the-art instance segmentation methods closer to non-experts by developing PlantSeg: a pipeline for volumetric segmentation of light microscopy images of biological tissues into cells. PlantSeg comes with a large repository of pre-trained models and delivers highly accurate results on a variety of samples and image modalities. We exemplify its usefulness to answer biological questions in several collaborative research projects. In the second part, I tackle the dense annotation bottleneck by introducing SPOCO, an instance segmentation method, which can be trained from just a few annotated objects. It demonstrates strong segmentation performance on challenging natural and biological benchmark datasets at a very reduced manual annotation cost and delivers state-of-the-art results on the CVPPP benchmark.

In summary, my contributions enable training of instance segmentation models with limited amounts of labeled data and make these methods more accessible for non-experts, speeding up the process of quantitative data analysis.