Visual Fine-grained Recognition

• Applicant: Professor Dr.-Ing. Joachim Denzler, since 1/2017
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: from 2015 to 2020
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 275610656

This project proposal deals with visual recognition of fine-grained object categories, an intersection topic between computer vision and machine learning. In general, research in visual recognition algorithms and models has received not only an enormous amount of attention by computer vision and machine learning researchers around the world, but shaped the path for many applications that are making it to everyday-life products. Examples include driver assistance systems with visual pedestrian detection, content-based image retrieval, or face localization in standard digital cameras and smartphones.Whereas research mainly focused on two very different situations; distinguishing between basic-level categories (category recognition) or recognizing specific instances (instance recognition), developing algorithms for automatically discriminating categories with only small subtle visual differences (fine-grained recognition) is a new challenge that just started in the last couple of years. In contrast to previous work in this area, we try to unify the approaches in fine-grained recognition with the ideas used for instance matching as well as category recognition. This will allow for exploiting and combining very basic concepts in all three branches, since in all of them, object part decompositions are directly or indirectly used for building feature representations and algorithms measuring the similarity between images. However, the differences between the areas are the different degrees of the variation of part appearances and the variation of joint geometrical configuration of part positions. Furthermore, the algorithms and models differ significantly ranging from orderless bag-of-visual-word models for category recognition to localized part features for fine-grained recognition tasks and finally to the matching of local features for instance matching. The aspects that we are planning to tackle in the proposed project will lead to improved recognition algorithms, which unify all three areas and automatically learn suitable part and feature representations from data itself. In particular, our proposed project aims at developing fine-grained recognition algorithms, where part-based models are automatically learned and adapted to new data and each test example. Therefore, the major objectives are:1. Unsupervised part constellation discovery and learning of part representations,2. Estimating exemplar-specific and local representations,3. Adaptation of part detection and classification models to new domains and tasks.The algorithms developed in the project will be important for several application branches, such as medical computer vision (e.g., distinguishing between different types of tissue), robotics ("Robot, please get me Bosch GSR professional screwdriver"), or biology ("What kind of bird is displayed in the image and what are its relevant features.") to give just a few examples.

• DFG Programme: Research Grants
• Ehemaliger Antragsteller: Professor Dr. Erik Rodner, until 12/2016