Reading Assignment: A Domain-Independent System for Sketch Recognition

Reference Information
Title: "A Domain-Independent System for Sketch Recognition"
Authors: Bo Yu and Shijie Cai
Citation: "A Domain-Independent System for Sketch Recognition", Bo Yu and Shijie Cai, Proceedings of the 1st International Conference on Computer Graphics and Interactive Techniques in Australasia and South East Asia, pp. 141-146, 2003.

Summary
This paper discussed a system for sketch recognition that accepts freehand sketches as input and performs recognition by using stroke approximation with low-level geometric features without the use of domain-specific knowledge. The output of the recognizer is a hierarchical structure of recognition information, designed to be easily used by high-level applications that the system can be embedded within. The system was designed with a list of ideal attributes for a sketch recognition system in mind. These attributes include the ability to draw naturally, to produce consistent recognition, to understand hierarchical relations, to predict the sketch during drawing, to provide an efficient and easy-to-use interface for the user, and be easily integrated into other applications.

The system has two main stages: stroke approximation and post-processing. Stroke approximation occurs during sketching, with each stroke as input as the stroke is completed. It uses recognition techniques to approximate the shape of the stroke as compared to a set of low-level geometric shapes (lines, arcs, circles, ellipses, and helixes). Stroke approximation includes vertex detection, line approximation with feature-area verification, curve approximation, and the handling of shapes with intersecting features. The post-processing stage is performed after the entire sketch is complete, and consists of using the data from the stroke approximation phase to create shape relations and complete recognition. It includes relation retrieval, cleanup (removing useless elements of the strokes and merging strokes as necessary), and finally, object recognition to recognize the basic objects that may occur within the gesture.

The user interface itself is designed to allow for creation, deletion, and modification of sketches. The modification feature is the most emphasized portion of the user interface. An evaluation was conducted by testing the system with users. It was determined that it was easy to use, provides useful information as output, and can be easily integrated into other applications as intended.

Thoughts
The idea of creating an easy-to-use recognition system for freehand drawing that can be easily integrated into other applications seems like it could be very useful. In addition, the attributes of an ideal, practical recognition system that were put forth could be very useful with regards to the future design of sketch recognition systems; however, I was curious as to where these properties came from. There is no source provided for the properties, so have they originated from research studies that have been conducted or are they simply the authors' opinions of the ideal attributes that a sketch recognition system should have? Since these are guidelines that the system is compared against throughout the paper, this could be an important distinction to make.

It was interesting to learn about methods for recognizing freehand sketches, as we have mostly focused on single-stroke sketches thus far. Also, the methods for vertex approximation mentioned the noise that is produced when using timing data, which is an interesting result to apply to the information that we learned previously of Sezgin's corner finding methods that used timing data.