Sketch recognition algorithms for comparing complex and unpredictable shapes

Martin Field, Stephanie Valentine, Julie Linsey, and Tracy Hammond. 2011. Sketch recognition algorithms for comparing complex and unpredictable shapes. In Proceedings of the Twenty-Second international joint conference on Artificial Intelligence - Volume Volume Three (IJCAI'11), Toby Walsh (Ed.), Vol. Volume Three. AAAI Press 2436-2441.


Mechanix is a free sketch recognition system focused on allowing students to solve free body diagram and static truss problems through sketching. The primary purpose is to provide quick feedback to students, while easing the burden of grading a high number of paper submissions for professors and TAs.

Mechanix uses an online process, in which after each stroke is drawn an attempt is made to combine it with others into a more complex shape. Mechanix employs Paleo sketch for low level recognition of strokes. Body identification is a primary task for Mechanix, which is solved by first identifying closed shapes and then entering a comparison phase. The comparison phase is unique, because only one template exists for comparison. An average of the Hausdorff distance, modified Hausdorff distance, and tanimoto coefficient is used with an acceptance coefficient of .65.

Mechanix also faces the issue of identifying trusses as drawn by students. Identifying shared edges is a key step in solving this recognition problem. Mechanix forms a graph with each line segment of the truss being a node and the edges being intersections. Using an elimination of connections technique, a shared edge can then be identified. For the final comparison both the sketch and adjacency graph are considered.

Mechanix represents an application of sketch recognition techniques to a specific domain. It is notable that new techniques were developed and tuned to handle the specific cases of the domain. The paper does not detail the feasibility or conflicts that might arise from attempting to include these techniques in a general recognizer. Further work on a generalized or extended domains approach would be interesting.