Modeling 3D animals from a side-view sketch - IRIT

1 Modeling 3D animals from a side-view sketchEven Entema,b, Loic Barthea, Marie-Paule Canib, Frederic Cordierc, Michiel van de PannedaIRIT - University of ToulousebUniversity of Grenoble-Alpes, CNRS (Laboratoire Jean Kuntzmann) and InriacUniversity of Haute AlsacedUniversity of British ColumbiaAbstractUsing 2D contour sketches as input is an attractive solution for easing the creation of 3D models. This paper tacklesthe problem of creating 3D models of animals from a single, side-view sketch . We use thea prioriassumptionsof smoothness and structural symmetry of the animal about the sagittal plane to inform the 3D reconstruction.

2 Ourcontributions include methods for identifying and inferring the contours of shape parts from the input sketch , a methodfor identifying the hierarchy of these structural parts including the detection of approximate symmetric pairs, and ahierarchical algorithm for positioning and blending these parts into a consistent 3D implicit-surface-based model. Wevalidate this pipeline by showing that a number of plausible animal shapes can be automatically constructed from asingle IntroductionWith the spread of 3D virtual environments and of 3 Dprinting technologies, many practitioners would like toauthor their own 3D shapes. Among them, animal mod-els including imaginary and fantastic ones are animportant category.

3 Being able to easily create and thenanimate animals would be an important step for generat-ing more lively virtual worlds. animals are also amongthe models that the general public, especially children,would typically like to sculpt and is currently no fast and easy method for creat-ing 3D models of animals . Unfortunately, getting datafrom 3D scans is much more difficult for animals thanfor humans, beginning with the obvious challenge of re-quiring an animal to stand still. In addition, such re-constructions are also limited to existing animals . Stan-dard 3D Modeling software, such as Autodesk s Mayaor Blender, as well as digital sculpting software such asPixologic-Zbrush, can be used for creating animals , buttheir complexity limits their use to experienced or pas-sionate users.

4 The use of 3D sculpting is possible, but isstill difficult: many people are not adept at sculpting an-imals using real clay and in this case, they are not likelyto perform much better in a virtual setting, even withan investment of time in the mastery of digital sculpt-ing interfaces. sketch -based Modeling systems, whichonly require users to sketch contours in 2D, are proba-bly the most intuitive and accessible class of , despite these advantages, they either requireusers to iteratively draw complex shapes part by part,using different viewpoints, or, alternatively, they requirean existing data-base of 3D work belongs to the category of sketch -basedmodeling methods and is the first to explore the cre-ation of a 3D animal model from a single, side-viewsketch.

5 We are motivated by the belief that many usersare capable of drawing a single sketch that depicts thePreprint submitted to Computer&GraphicsSeptember 30, 2014contour lines and the internal silhouettes of an animal ,such as shown in the top-left of Figure 1. If needed,users can use a background drawing or a photo as aguide. The process of inferring 3D geometry from the2D sketch necessitates the use of relevant assumptionsin order to be tractable, and in our context of modelinganimal forms, we shall assume smoothness of the result-ing shape as well as the presence of structural symme-tries. Several further moderate constraints include: (a)restricting ourself to non self-overlapping limbs in thesketch; (b) requiring the user to draw both contours forpairs of symmetric limbs; and (c) ignoring the recon-struction of repetitive details scattered on the surface,such as scales.

6 With these assumptions and constraintsin place, the method we develop is capable of automat-ically converting an input vectorized sketch into a 3 Dmodel. The total processing time is less than one sec-ond, effectively enabling one to create new animal mod-els in only the time required to sketch that in this work, we only tackle the creationof the volumetric shape parts of an animal and that wedo not consider the surface components that should beused for ears or scales. The ears we reconstruct arealso therefore interpreted as volumes. We are not ableto reconstruct large flat parts such as wings. In addi-tion, we reconstruct limbs in a symmetric fashion, evenwhen they were drawn in arbitrary postures in the inputsketch.

7 The construction of symmetric 3D models isusually desirable, as it ensures that left and right limbshave identical dimensions. To achieve a desired non-symmetric posture, the 3D model can be deformed, ei-ther by using an animation skeleton and the associatedskinning weights, or by directly articulating the implicitsurface s processing pipeline for creating 3D animals froma sketch is summarized in Section 3. It consists of threemain steps, which also correspond to our three technicalcontributions:1. the identification of the animal s foreground struc-tural parts in the sketch , with completion of theparts that are not explicitly bounded, such as thetop of the legs (see the teaser figure)2.

8 The generation of a hierarchical graph of depths forthe structural parts, using the complete results andbased on an algorithm for detecting the portions ofthe sketch that correspond to symmetrical parts ofthe animal ;3. 3D reconstruction based on a specific choice ofimplicit surface, scale invariant integral surfaces,which enables us to accurately reconstruct shapeparts from their medial axis in the 2D sketch ,and to seamlessly blend them into a single Related Work2D sketches only represent the contours, silhouettesand main features of an object. Converting them intoa 3D model therefore requires resolving indetermina-cies and inferring a large amount of missing data.

9 Fourstrategies are commonly used to do so, each of whichis based on a different level of hypotheses ora prioriknowledge of the shape being modeled (see [1] and [2]for detailed surveys):Iterative methodsenable the user to build general mod-els part by part, by iteratively adding new shape com-ponents from different viewpoints, the Teddy sys-tem [3]. These methods make the hypothesis that thefinal shape is a combination of parts that all have planarsilhouettes from a given viewpoint, and can therefore beinflated from closed planar contours. animals belong tothis category but in practice these methods still requirepractice and time in order to achieving a convincing re-sult.

10 Other iterative methods have been developed us-ing alternatives to inflation for geometric include the use of primitives such as generalizedcylinders and ellipsoids, as in [4], and methods basedon implicit surfaces, as in [5] or [6].Shape matchingapproaches match the user sketchwith silhouettes, or parts of silhouettes of predefined3D models, possibly enabling some deformation. Thismethod was successfully applied to organic shapes suchas humans or animals [7], as well as for technical mod-els [8]. However, they require a template example of thegiven general class of animal which imposes a restric-tion on the family of sketches that can be used.