Graph Active Semi-supervised Semantic segmentation for event-based vision
Semantic segmentation is an important topic in computer vision. The task is to classify every pixel of an image according to what is being represented. Semantic segmentation has several applications in computer vision including autonomous driving, medical image diagnostics, among others. Many researchers have applied successfully Convolutional Neural Networks (CNNs) to the problem of semantic segmentation. However, CNNs are very complex models that require a lot of data to avoid the over-fitting problem. Indeed, there are no general answers in the literature about the sample complexity required in the deep learning regimen. The over-fitting problem is especially sensitive when we do not have a lot of labeled information. For example, the study of event cameras is a promising field in autonomous driving due to several advantages that these devices have over conventional cameras. However, this field does not have the huge amount of labeled data that exists in conventional RGB cameras, limiting the application of deep CNNs. Semi-supervised learning is a research field of increasing interest in machine learning. This field can potentially solve the problem of labeled-data requirements since its importance lies in the availability of large amounts of unlabeled data, and the interest to exploit it. The goal is to improve the quality of predictions and inference in downstream applications. To this end, an effective semi-supervised learning method must learn from the labels and structure of the datasets. Similarly, active semi-supervised learning tries to achieve the best performance using labeled and unlabeled data. However, in this case, the model can select the best-labeled samples beforehand to classify the entire dataset. In this work, we introduce a new algorithm for active semi-supervised semantic segmentation. Our algorithm is inspired by the sampling of Graph Signal Processing (GSP) and Graph Convolutional Networks (GCNs). Our algorithm will be composed of 1) regions proposal, 2) features extraction, 3) inference of the underlying graph, 4) sampling of graph signals, and 5) a GCN for semi-supervised learning.