Reading #10: Graphical Input Through Machine Recognition of Sketches (1976)

by Chrisopher F. Herot (paper)

Comments: Jonathan

This is an early sketch recognition system aiming to allow sketch input to computer programs. It contains a system called HUNCH that is used to recognize primitive sketches. It uses speed only to detect corners in a sketch. Curves were viewed "to be a special case of corners," and were modeled using b-splines. Speed was also used to determine how "careful" the user was, for instance faster strokes were less careful. This was used to help identify curves and decide whether to draw them as b-splines or just corners. The programs used to convert sketches to straight segments and curves were called STRAIT and CURVIT, respectively.

STRAIT and CURVIT did not always create the same interpretation as humans. They seemed to be user-dependent, as sketches were interpreted better for some users than for others. An improved method of straight segmenting was implemented, called STRAIN. It used a function of speed to determine which line endpoints to join (instead of a fixed distance, which is what STRAIT did).

Programs were also developed to detect overtraced lines and turn 2D sketches into 3D.

The paper discusses the importance of context in sketch recognition systems. The HUNCH system does not use context in its recognition schemes. For example, all its subroutines are always called in a fixed sequence and always perform recognition in the same way. However, similar strokes will probably be interpreted in different ways given their context.

The paper discusses an interactive system and examines the hierarchical structures of recognized sketches. It discusses various ways to tune the algorithms to work for a "truly interactive system."

__________

This is an early approach to sketch recognition, and it asks many questions as well as answers some. It can be rather dry and boring, but it does bring up some questions that are still relevant today and whose solutions can still be improved on.

For example, the latching problem: when should close endpoints be merged? It depends on the context, which is another contemporary issue. While I was working on my truss recognizer, I dealt with the latching problem when merging close line endpoints to form truss nodes. I used a distance computed based on stroke lengths, but that was obviously a poor choice.

This paper begins to explore machine learning techniques for interpreting sketches. It introduces many questions and proposes possible solutions to those questions by hypothesizing extensions to an existing primitive corner finding and beautification system. I believe the authors asked good and relevant questions, as we are still asking ourselves how to solve many of the same problems in better ways.