ECCV 2020 Daily - Wednesday

2 Poster Presentation 16 Razieh Kaviani Baghbaderani is a third-year PhD student in electrical engineering at the University of Tennessee, Knoxville. Ying Qu is a research associate at the University of Tennessee, Knoxville, in the same lab as Razieh, with a focus on computer vision and remote sensing. They speak to us ahead of their poster tomorrow. Razieh and Ying’s work is about open- set recognition for satellite imagery. It was inspired by a research project they worked on last year exploring land cover classification on satellite imagery. Land cover classification is an important step towards analyzing theEarth’s surfaceand means assigning pre-defined material categories to each pixel of the satellite imagery. The team have a large collection of satellite imagery fromall over theworld, covering vast areas and a wide variety of materials from different cities and countries. Satellite imagery is important for gaining a better understanding of key global issues such as climate change and urban development . Land cover classification has been deeply involved in human activities like ecological science, precision agriculture, and military applications. Traditional methods for land cover classification have closed-set settings, meaning they assume the training samples in the training dataset will cover all the classes in the testing dataset. In reality, that is impossible. Themotivation for this project is to figure out how to identify the unknown classes to help human annotators to collect training samples from them. Razieh tells us they need comprehensive training data: “ To get this training data, we need annotators who work manually, which is hard work and very time consuming. Our paper proposes a system that helps them by automatically identifying if an image has novel categories that have not been seen before. It can detect and label new categories to add new samples to already available training data and actively improve and enhance the model as it goes.” Representative-Discriminative Learning for Open- set Land Cover Classification of Satellite Imagery DAILY W e d n e s d a y