Change Detection Tool Based on GSV to Help DNNs Training

摘要

We present a system to carry out the automatic detection of structural changes through a Deconvolutional Neural Network (DNN) in images synthesized from panoramas provided by an online and open source map tool, Google Street View (GSV). Our approach is motivated by the need of more efficient and frequent updates on large-scale maps for autonomous driving applications. To train and evaluate our DNN we build a geolocation database, an order of magnitude larger than other existing datasets, based on pairs of images and their corresponding ground truth that shows changes detection over time. A tool has been implemented to guide manual annotation of changes using panoramas all over the world. The tool chains the panoramas and depth maps creation, the image synthesis and the labelling synthesized images generating their groundtruth. Finally, a DNN has been trained to automatically detect changes validating our methodology by using the obtained dataset, yielding better results that other state-of-the-art approaches.