1A review of: A Sketch-Based Interface for the Design and Analysis of Simple Vibratory Mechanical SystemsChris EllisOctober 31, 2007Sketch-based Design of Vibratory Mechanical Systems Rapidly construct diagrams of mechanical systems Simulate the motion of the system Graphs the results2Sketch-based Interface Can draw masses, springs, dampers, forces, and grounds Can use any number of strokes in any order Can process and interpret strokes as simulated vibratory systemSketch-based Interface (cont.) Labels recognized symbols User can tap symbols and change parameters (also sketch-based) User can toggle between ink and “beautified” mode3Mark-group-recognize Pull out easy-to-recognize symbols Cluster remaining strokes into symbols Perform more difficult recognition on remaining strokesMarking Step Extract masses Determine closed loop by finding minimum endpoint distance to close stroke chain Threshold by 10% of total stroke length Limit 5 strokes Allows for arbitrary stroke order, direction, and shape4Marking Step (cont.) Extract grounds Test if separation between pair of most distant consecutive strokes < twice the average separation Test if hashes all point to same quadrant (NW, NE, SW, SE) Skeleton length w/in 5% of length of line from first hash to last Find minimum of 4 hatch strokes meeting above criteria Add hatches until a stroke breaks the patternGrouping Step Still need to identify springs, dampers, and forces Concerns: Can be arbitrarily close Can be arbitrary size, shape, number of strokes Do not know number of clusters5Grouping Step (cont.) Uses domain specific knowledge Clusters are spatially distinct Clusters do not overlap Use agglomerative hierarchical clustering Each point initially belongs to its own group Merge the two nearest-neighbor groups until dissimilarity score jumps drasticallyRecognition Step For each cluster Find segment points (speed/angle cusps) Use least squares to fit arcs and lines to ink between the segment points Extract nine features (e.g. # of strokes, # line segments, # arc segments, # L intersections, # X intersections, # T intersections, # pairs of parallel lines, # pairs of perpendicular lines, avg distance between endpoints) Classify6Recognition Step (cont.) Classification Assumptions Assumes training features are distributed normally (Gaussian) Assumes features are statistically independent Neither assumption is true, but good enough Use Naïve Bayesian ClassifierConnectivity Similar to Grouping Step, connects nearest components together Euclidian distance bounding box center for mass endpoints for others Additional constraints Springs or dampers connected to exactly one mass or ground Mass or ground may have any number of springs or dampers7Finally… Results of the recognition used to construct motion equations Equations described as matrices and vectors Matrices and Vectors passed to Matlab Displacement vector returnedD = {d0(t),d1(t),…dm-1(t)}Evaluation Testers generally had little to no experience with Tablet PCs Given ~30s of practice time No explanation of system functions8Evaluation (cont.) Worked correctly for majority of subjects Errors from unexpected input Symbols too close Symbols drawn strangely False positives for massesThe EndOr is it?(Yes it
View Full Document
Unlocking...