Interpreting Physical Sketches
as Architectural Models

Barbara Cutler and Joshua Nasman
To appear in Advances in Architectural Geometry 2010

paper     conference website     publisher website

Abstract: We present an algorithm for the automatic interpretation of a rough architectural sketch as a consistent 3D digital model. We compare our results to the designer's intended geometry. We further validate the algorithm by studying the variations in possible interpretations made by other humans for a set of relatively ambiguous sketches. In our system, the user sketches an architectural design by arranging small-scale physical wall modules and simple markers for windows on a table. These color-coded elements are captured by a camera mounted above the scene and recognized using computer vision techniques. The architectural design is automatically inferred from this rough physical sketch transforming it into a consistent and manifold 3D triangle mesh representation. The resulting digital model is amenable to numerous building simulations including lighting, acoustics, heating/cooling, and structural analysis.


Figure 1: In our physical architectural design environment a) users gather around a table and construct b) a small-scale mockup of a design from a collection of wall elements and marker tokens. A camera above the table c) captures the layout of elements on the table. Our sketch interpretation algorithm processes these elements to construct a consistent and watertight triangular mesh of the implied architectural design (ceiling removed for visualization).


Figure 2: Intended collinearity can be ambiguous: a) detected primitives, b) annotation by the original designer, and c-f) annotation by other users.


Figure 3: Dividing space into cells by extending tangents and connecting elements: a) detected primitives, b) annotation by original designer, c) wall chains, d) zones defined by the wall chains, e) enclosure, and f) our automatic interpretation.


Figure 4: Designs with non convex boundaries may not be accurately extracted with a simple threshold on the average point or cell enclosure. By minimizing the lengths of unused wall and extra inferred wall necessary to enclose the interior zones we correctly interpret these complex designs.


Figure 5: Challenging examples of designs with multiple rooms. The examples in the bottom row are somewhat ambiguous and have multiple reasonable interpretations for the passageways between rooms.


Figure 6: If the design contains a interior room with no exterior walls, this space may be intended as a courtyard space, uncovered by a roof. The system will require extra information from the designer if the default interpretation does not match his/her intentions.


Figure 7: Domain-specific knowledge may be necessary to correctly interpret sketches that hint at common architectural forms, such as the cruciform used in church floor plans (top row) or to recognize an entrance portico (bottom row). Despite the potential for ambiguity, b-e) most users' interpretations matched a) the original designer's intention. Our automatic sketch interpretation results are shown in f).


Figure 8: Some ambiguous designs: a) the original designer's annotation, b-e) annotation by other users, f) our automatic sketch interpretation results.


Figure 9: A sampling of the variety of collected physical sketch geometry: the original designer's annotation, and our automated interpretation of the interior/exterior space.

See also:
RPI Computer Graphics Group
Research Project: Architectural Daylighting

This material is based upon work supported by the National Science Foundation under Grant No. CMMI 0841319 and Grant No. IIS 0845401. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.