An Object Detector Trained on Line Drawings

We investigate line drawings as an alternative method to train an object detector. While traditional approaches rely on learning an object class from example images, we make use of abstraction provided by the artist. Our object class of choice are cats, posing a challenging detection problem by a variety of poses and appearances. In order to deal with this flexibility, and similar to part-based models like poselets, we decompose each line drawing into smaller units. HOG feature descriptors are then used in multiscale sliding window detection Using the PASCAL Visual Object Classes framework for testing, we determine some parts with promising detection performance. Based on example image patches obtained during this procedure we train support vector machine classifiers in order to improve our initial detectors. While some part detectors already perform quite well, SVM training displays mixed results possible reasons of which we discuss in detail.