Abstract by Taylor Sorensen
Pose Alignment in Computer Vision
An important problem in computer vision is pose-alignment, which is the identification of several key points of an object in an image. For example, on a person, we might want to identify the head, feet, hands, and joints. However, this is a non-trivial problem, since clothes and physical features vary from person to person. Being able to accurately map these points allows for much more accurate fine-grained visual categorization (FGVC), which focuses on identifying on a smaller level: not only if a picture is of a bird or a dog, but the species of the animal. Many of these differentiating factors can be more easily spotted by localizing and comparing them between poses. Our research focuses on being able to accurately train a model to recognize animals’ pose, and thereby be able to differentiate between similar species. To do so, we are collecting hours of animal videos, on which people then map out similar points. Using visual tracking software, we extrapolate these points on select frames to the other frames in the video, creating a rich dataset on which we can train our model.