Bibliography:
Brandon Paulson and Tracy Hammond. PaleoSketch: Accurate Primitive Sketch Recognition and Beautification. Proceedings of the 13th international conference on Intelligent user interfaces. 2008
Summary:
The paper presents a system - PaleoSketch, that can accurately identify eight low level primitive shapes (line, polyline, ellipse, circle, arc, curve, spiral and helix) and hierarchically construct higher-level shapes based on geometric constraints defined by the user. Additionally, it also beautifies the sketch, thus providing early feedback.
The general architecture of the system consists of three steps: Pre-recognition, Shape Tests and Hierarchical interpretation. The pre-recognition, in addition to resampling, computing corners, speed and direction metrics, calculates two new features: Normalized distance between direction extremes (NDDE) and Direction Change ratio (DCR). These metrics are at the heart of the various shape tests. NDDE is computed by taking the distance between the points with the highest and lowest slopes. (change in y / change in x) This is normalized by the total stroke length. DCR is calculated by taking the max change in direction and dividing by average change in direction. Before performing the various shape tests, the system also determines if the shape is overtraced and is closed.
A shape is classified as complex when none of the shape tests pass. The final part of the system, the hierarchy function, sorts the results in order of best fit. The system was found to perform well in the given dataset and is comparable to the best low-level recognizers, that don't handle so many shapes.
Discussion:
PaleoSketch is a great example of a geometry based recognizer, and shows how far a constraint based sketch recognition system can go. It was nice to see simple metrics such as NDDE and DCR to be very effective in identifying primitive shapes. I wonder how the authors arrived at the various thresholds used in the paper. (Mainly during preprocessing, when checking for overtracing and if the figure is closed), as those seem crucial to the performance of the shape tests.
Brandon Paulson and Tracy Hammond. PaleoSketch: Accurate Primitive Sketch Recognition and Beautification. Proceedings of the 13th international conference on Intelligent user interfaces. 2008
Summary:
The paper presents a system - PaleoSketch, that can accurately identify eight low level primitive shapes (line, polyline, ellipse, circle, arc, curve, spiral and helix) and hierarchically construct higher-level shapes based on geometric constraints defined by the user. Additionally, it also beautifies the sketch, thus providing early feedback.
The general architecture of the system consists of three steps: Pre-recognition, Shape Tests and Hierarchical interpretation. The pre-recognition, in addition to resampling, computing corners, speed and direction metrics, calculates two new features: Normalized distance between direction extremes (NDDE) and Direction Change ratio (DCR). These metrics are at the heart of the various shape tests. NDDE is computed by taking the distance between the points with the highest and lowest slopes. (change in y / change in x) This is normalized by the total stroke length. DCR is calculated by taking the max change in direction and dividing by average change in direction. Before performing the various shape tests, the system also determines if the shape is overtraced and is closed.
A shape is classified as complex when none of the shape tests pass. The final part of the system, the hierarchy function, sorts the results in order of best fit. The system was found to perform well in the given dataset and is comparable to the best low-level recognizers, that don't handle so many shapes.
Discussion:
PaleoSketch is a great example of a geometry based recognizer, and shows how far a constraint based sketch recognition system can go. It was nice to see simple metrics such as NDDE and DCR to be very effective in identifying primitive shapes. I wonder how the authors arrived at the various thresholds used in the paper. (Mainly during preprocessing, when checking for overtracing and if the figure is closed), as those seem crucial to the performance of the shape tests.
No comments:
Post a Comment