CAD-based Systems

A number of papers approach the problem of model generation from an en- tirely different direction. CAD-based systems attempt to create solid objects from technical or architectural drawings, often by calculating depth as op- posed to the common use of implied depth in artist-driven systems. The most important aspect common to most CAD-based construction systems is that little or no user interaction is required, and input imagery is often only single-view.

One of the most successful attempts at closing the gap between 3D gener- ation of solid and organic surfaces is ‘3D Sketch-Based Model Reconstruction and Rendering’ by Mitani, J et al. [123]. Their paper proposes a system that takes a user-drawn CAD-style image in 2D and attempts to find the best- fit bounding boxes, including warping of the surface to better fit non-linear shapes. While the results show impressive accuracy, this comes at the cost of a limited input range. In section 4.1 they outline a number of constraints that must be fulfilled for the system to work, including having mirror sym- metry and specified planar faces. The edge graph used in the paper can be easily extracted due to the sketch-based input method, whereby the image construction is captured during the drawing phase on a graphics tablet.

The corner and edge detection allow a base 3d model to be created with planar faces. Curve fitting and reconstruction is then used to compare the drawn line between two corners with the reconstructed edge. The curved line is projected into 3d space using the reconstructed bounds and the camera properties, and is then used to modify the curvature of the faces it bounds. This two-step process of rough cage creation followed by detail adjustments results in models that appear correct at both macro and micro scale.

Few other papers in this group attempt to model curved surfaces. Indeed, the technique outlined in Reconstructing Polyhedral Swept Volumes from a Single-View Sketch [159] requires the single-view input sketch to contain well- defined corners. The input image is also assumed to contain one profile view, and show the tail end of an extruded solid. Orthogonal and non-orthogonal corners are recognised in the image, and used to find the plane of the profile and the orthogonal direction of the extrusion. These can then be used to create the 3D model. The system relies on the input image being a relatively correct geometric sketch.

One of the primary constraints and advantages of our system is that it uses a single input view. Suh [159] outlines a method of determining orientation given only a single view, and while this only applies to CAD- style mechanical objects, any input image with potential right-angled corners could be analysed and use this method to determine the initial orientation. Three-quarter views are common among concept artwork, and being able to un-distort the initial image could result in higher quality models.

Another system that is similar in technique to those outlined above is the Smart Sketch System for 3D Reconstruction Based Modeling by Naya et al. [129]. The paper outlines a freehand CAD system that automatically straightens freehand input lines and aligns them in viewspace with a prede- fined projection. A second input step is then performed with the surfaces being sketched on top of the reference lines. The program automatically fills these in to create the model. Although the view angle is unconstrained, ex- amples show that the system works from a single-view input sketch. This two-step mesh generation algorithm appears to afford better control over how the surface is created, as guidelines and the reference frame can be set up first and then adjusted to fix alignment and perspective problems before any ge- ometry is created. However, to implement this using concept art would also require a 2-step input, perhaps in the form of a sketched wireframe followed by the painted concept image.

The majority of 3D reconstruction techniques that fall within this cate- gory have trouble reconciling the implied accuracy of technical drawings with the low accuracy of hand-drawn sketches. Fen et al. [59] use vanishing points, detected by tracing any straight lines in the image, to determine the focal

length and perspective of a drawing. This is then used to identify corners and estimate depth data to recover the 3D object. This worked well for accurate drawings, however for inaccurate drawings such as sketches the algorithm had greater trouble calculating the vanishing points and focal length.

In opposition to artist-assisted sketching interfaces, CAD-based genera- tion systems rarely rely on time-series data, nor multi-view data input. Some of these techniques could prove applicable when generating character mod- els, such as in the case of robotic or highly stylised drawings. However this situation is not as common as organic, arbitrary shaped characters and the principles behind polygonal mesh synthesis from sketches is fundamentally different from rounded mesh synthesis. In addition, the concept artwork used as a template does not often have a rigid reference frame and may not be strictly orthogonal, perspective, or isometric. CAD-based single view recon- struction relies upon accurate adherence to this outside frame of reference.