Bibliography:
Brandon Paulson, Pankaj Rajan, Pedro Davalos, Ricardo Gutierrez-Osuna, and Tracy Hammond. What!?! No Rubine Features?: Using Geometric-based Features to Produce Normalized Confidence Values for Sketch Recognition. VL/HCC Workshop: Sketch Tools for Diagramming. 2008.
Summary:
This paper presents a hybrid sketch recognition system, for recognizing lower-level primitives that combines features from both gesture and geometric recognition systems. The system aims at making the recognition user independent and allow for interpretation with normalized confidence values.
The system uses a quadratic classifier that contains both geometric and gesture features. The feature set consists of 44 features, 31 geometric and 13 of rubine's gesture based features. Feature subset selection was done using a greedy, sequential feed forward (SFS) technique. Using this subset of features, the accuracy achieved was comparable to that of the PaleoSketch system.
The system is easier to code and performs faster classification when compared to the PaleoSketch system as fewer features need to be compared. 14 of the 31 geometric features were present in the optimal subset, which included NDDE and DCR, (See PaleoSketch), least squares line, major-minor axis ratio. Only total-rotation was chosen from the gesture based features in the optimal subset, which is mainly because the data was chosen such that it was user independent.
Discussion:
It was nice to see that geometric constraints play an important role in user independent recognition systems. While gesture recognition has its place in the field of sketch recognition, I see geometric recognition to be a key player in bringing sketch recognition systems to main stream mobile platforms. Thinking of app developers who'll be building apps on these gesture based devices, I think it makes more sense for them to use recognition engines that are user independent.
Brandon Paulson, Pankaj Rajan, Pedro Davalos, Ricardo Gutierrez-Osuna, and Tracy Hammond. What!?! No Rubine Features?: Using Geometric-based Features to Produce Normalized Confidence Values for Sketch Recognition. VL/HCC Workshop: Sketch Tools for Diagramming. 2008.
Summary:
This paper presents a hybrid sketch recognition system, for recognizing lower-level primitives that combines features from both gesture and geometric recognition systems. The system aims at making the recognition user independent and allow for interpretation with normalized confidence values.
The system uses a quadratic classifier that contains both geometric and gesture features. The feature set consists of 44 features, 31 geometric and 13 of rubine's gesture based features. Feature subset selection was done using a greedy, sequential feed forward (SFS) technique. Using this subset of features, the accuracy achieved was comparable to that of the PaleoSketch system.
The system is easier to code and performs faster classification when compared to the PaleoSketch system as fewer features need to be compared. 14 of the 31 geometric features were present in the optimal subset, which included NDDE and DCR, (See PaleoSketch), least squares line, major-minor axis ratio. Only total-rotation was chosen from the gesture based features in the optimal subset, which is mainly because the data was chosen such that it was user independent.
Discussion:
It was nice to see that geometric constraints play an important role in user independent recognition systems. While gesture recognition has its place in the field of sketch recognition, I see geometric recognition to be a key player in bringing sketch recognition systems to main stream mobile platforms. Thinking of app developers who'll be building apps on these gesture based devices, I think it makes more sense for them to use recognition engines that are user independent.
No comments:
Post a Comment