Learning Shape-Proportion Relationships from Labeled Humanoid Cartoons Md. Tanvirul Islam School Of Computing National University of Singapore tanvirulbd@gmail.com Yong Peng Why Faculty Of Arts and Social Sciences National University of Singapore psywyp@nus.edu.sg Golam Ashraf School Of Computing National University of Singapore gashraf@nus.edu.sg Abstract- Character design artists typically use shape, pose and proportion as the first design layer to express role, physicality and personality traits. Inspired by this we approach the problem of automatic character synthesis by attempting to learn relations among the body-shape, proportions, pose, and trait labels from finished art. In our prior work [13], we have designed an online game framework to collect and analyze perception data on hundreds of humanoid characters. We clustered the labels and then established a relationship between the body shapes and the pose-proportion feature space. In this paper, we extend the work to explore partial shape synthesis of a character’s torso and abdomen, given an input pose and proportion feature set. This paves the way for fully automatic character synthesis from labels. This is an improvement of our prior work, which addressed only shape classification. Keywords: Neural networks, Shape-proportion learning, perception modeling, shape synthesis I. INTRODUCTION One of the first things children learn is to manipulate and express with primitive shapes. Even as adults, we naturally tend to decompose complex compositions into primitives. Basic shapes like triangles, circles and squares are so well understood, that even a textual/ verbal description of structures in terms of these shapes elicits a natural visualization in our brain. Basic shapes play an important role in design drafts. For example, artists use shape scaffolding to pre-visualize the final form, using basic shapes to represent each component or part. Apart from establishing the volume and mass distribution of the figure, these shapes may also help portray a certain personality, as is widely seen in stylized cartoon drawings. For example, in Pixar’s recent animated feature titled UP, the main protagonist had distinctively square features to highlight his cooped-in life. The square features were amplified by contrasting with a large round nose, as well as distinctly rounded supporting characters. Depending on the art style, primitive shapes may become less apparent with the addition of details; e.g. clothes, accessories, and hair for humanoid figures. We use a vector shape representation that allows consistent blends between square, circle and triangle primitive shapes to express the shapedness of each body-part [12]. We will explain later how this vector primitive based representation of shapes plays a key role in our learning and auto synthesis framework. The main contribution of this paper is a neural network based expert system that takes as input the body pose and proportions of particular parts and then suggests the vector shape of body parts according to a required character trait. The synthesized shapes will provide a suggestive scaffold for the artists to add details on. By putting the details like clothes, accessories, and hair and drawing the figures in a fuller form the artist gets a humanoid cartoon with a particular character trait. This allows even an amateur or a newbie to create cartoons according to the desired role, physicality, and personality. We explain the methodology for shape learning and synthesis. Though this can be applied to create independent models for the full body, we demonstrate results for two body parts that tend to show a lot of shape variety, namely the torso and abdomen. We demonstrate the synthesis results for three distinct body types: fat, thin, and muscular.