Abstract by Connor Anderson
Fine Grained Recognition Using Planar Semantic Parts
Fine-grained Recognition poses the task of determining which, among many visually similar categories (e.g. different bird species), is the correct one for an observed object. Despite the diverse range of approaches that have proven successful for some object domains, fine-grained recognition remains a challenging and unsolved problem for others. Butterflies are just such a domain -- traditional methods, which use keypoints to represent parts, are inherently ill-suited to model a butterfly's approximately planar wings. To directly tackle this difficulty, an innovative framework is proposed, focused on representing objects with planar surfaces and leveraging semantic part segmentation to find these planar parts. Using a novel pose alignment model, the location and orientation of each such part is predicted and its respective appearance features are mapped into a pose-normalized space for recognition. A comparison with state of the art methods is included, showing improved recognition accuracy.