Semantic Segmentation in Geospatial Computer Vision

点击翻译按钮获取中文摘要

Recently there have been tremendous advances in computer vision due to the advent of deep learning. Human-like performance has already been achieved for cognitive tasks involving visual and spatial processing. However, this high performance is strongly dependent on the number of training examples in the dataset. To address this, many online competitions are offering large benchmark datasets for training. The creation of training datasets of images or video is a costly and labour-intensive process. It requires a significant number of people to label the data and ensure its correctness and completeness manually. This is exacerbated by the fact that a dataset can typically only be used for a single classification task. This is especially the case in the geospatial domain working with remote sensor data where most datasets are for building/non-building classification. Despite the difficulties involved in the creation, researchers rely on large datasets for training classifiers to assist in solving more difficult problems such as reconstruction. Reconstructing large-scale urban areas is an inherently complex problem that involves several vision tasks. The first step is semantic segmentation(*), where the objective is to label each pixel into an urban feature type, e.g., building, road, tree, vegetation, cars, clutter etc. Next, the pixels are clustered based on their labels into contiguous groups corresponding to instances of the urban features they represent. Finally, the reconstruction is performed on each cluster, where a customized algorithm is applied according to the urban feature type corresponding to the cluster. Hence, as it is evident, to achieve a complete urban-area reconstruction, one must first address the problems relating to semantic segmentation(*). This research program builds upon our most recent research outcomes in creating large-scale realistic virtual environments and focuses on addressing some of the significant challenges identified so far. Specifically, this DG will investigate the following two research objectives: 1.Few-shot semantic segmentation. The objective is to investigate network architectures and training paradigms which enable network training using only a minimal set of training examples. 2.Interpretability. The objective is to investigate methods for analyzing and interpreting what the network is learning internally with the goal of re-purposing the abundant pre-trained semantic segmentation networks on auxiliary tasks relating to their primary task, without further training or fine-tuning. This research program is expected to make substantial contributions to the solution of complex problems of high practical relevance to the field of computer vision. (*)Semantic segmentation: each *pixel* has its class label; Classification: the *image* has a single class label. Acronyms used to describe the progress of PhD students/candidates: CE: Comprehensive Exam RP: Research Proposal DS: Doctoral Seminar