Abstract

Monitoring species diversity or groups of individuals is important for many ecological studies, and algorithms of computer vision and machine learning are a great support for handling the huge amount of data that is recorded by camera traps. However, no matter how large an annotated dataset is for training a recognition system before deployment, a fixed model will be suboptimal for long-term monitoring due to several changes that naturally happen over time, e.g., shifts of the input data distributions and even a varying number of species or individuals that need to be identified. Therefore, we argue that recognition systems are required to adapt to new situations and conditions during the application. For this reason, the concept of lifelong learning is presented in this talk, which combines several aspects of continuous learning, including incremental learning of model parameters, active learning to incorporate expert feedback, and novelty detection to spot new classes or something unusual like anomalies. In addition, further research activities of the Computer Vision Group Jena are briefly summarized.

Biography

Paul Bodesheim received a diploma in computer science (Dipl.-Inf.) as well as a PhD in computer science (Dr.-Ing.) from the Friedrich Schiller University Jena (Germany) in 2011 and 2017, respectively. As a PhD student in the Computer Vision Group of the Friedrich Schiller University Jena, he mainly focused on novelty detection in visual object recognition and his dissertation is entitled “Discovering unknown visual objects with novelty detection techniques”. Afterwards, he joined the Max Planck Institute for Biogeochemistry Jena (Germany) as a postdoc, working for the EU Horizon 2020 project “BACI: Towards a Biosphere Atmosphere Change Index”. His research topic in this project has been the application of machine learning algorithms for predicting Earth observation data, in particular for upscaling site-level measurements to produce spatially resolved maps. In 2018, he returned to the Computer Vision Group of the Friedrich Schiller University Jena as a postdoctoral researcher, where he is a team leader for “Computer Vision and Machine Learning” since June 2020. His main research interests are visual object recognition, learning from small and imbalanced data, novelty detection and open set recognition, active learning and lifelong learning, as well as fine-grained recognition and its applications in biodiversity research, including the identification of plants, mammals, birds, and insects.

Further information on Paul's research is available on his website