Detecting garment and its landmarks

• Autor(a) principal: Daniel Fernandes Gomes
• Data de Publicação: 2017
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/107701
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

Garment folding is a task happening daily at our homes, retail and industry. When put into numbers, over a lifetime people spend on average 375 days performing this chore, while employees at a store may fold the same shirt 119 times per day. Despite the associated repetitive characteristics of this task, its automation is still far from being achieved mainly due to the large number of possible configurations that a crumpled piece of clothing may assume. In general, highly deformable objects still present large challenges for both fields of Robotics and Computer Vision. We attempt to offer a contribution to the garment folding automation by addressing the recognition of clothing pieces without much constraints on their pose or wrinkling, mimicking a most realistic scenario as possible. Such capability would enable a folding robot to choose and adapt its execution plan to the current clothing category. Because the considered problem revolves around clothe recognition, this work may also be of the interest of many other clothe related software applications such as recommendation systems existing on e.g., online e-commerce platforms, or intelligent surveillance setups that require tracking of people by their clothing description. Some work has been produced using Machine Learning techniques that, in general, consist on extracting a set of engineered features from the source image and then applying classification algorithms (e.g., Support Vector Machines) to find the associated clothing category and or pose. With the recent success of Convolutional Neural Networks, where features extraction is incorporated in the learning process, on the object classification problem, these have been preferred in favor of the previous pipelines. We apply Deep Learning techniques on images containing a single piece of clothing in a flat, wrinkled and semi-folded pose, existing on a clean background with the goal of classify and localize each piece. Furthermore, its relevant landmarks (shoulders, legs, crotch, etc) are equally treated. We train and evaluate our solution using the datasets produced by CTU at the CloPeMa project.