Reading #5

Comment:
chris aikens


Summary:



In this paper, the author has introduced the 1$ recognizer, as the name indicates, this algorithm is simple, easy to implement and can be docked into other system easily.

Its algorithm can be described in four steps,

1. Resample the Point Path, to make gesture paths directly comparable even at different movement speeds.
2. Rotate once based on the “Indicative Angle”, where “indicative angle” is defined as the angle formed between the centroid of the gesture (x¯,y¯) and the gesture’s first point.
3. Scale and Translate, during which the gesture is scaled and translated to a reference square and point.
4. Find the Optimal Angle for the Best Score, after all the previous steps are done for all templates, a candidate C is compared to each stored template T to find the average distance, which is defined as the lease path-distance. The template Ti with the least path-distance to C is the result of the recognition.

The $1 recognizer is limited to single stroke, and the author states an accuracy more than 99% accuracy, as well as a comparison with Rubines algorithm and Dynamic Time Warping.


Discussion:

Feature based vs Template based: Rubines and many other researchers formed the gesture recognition as a feature-based recognition process, as far as the fractures can be easily computed (which is usually true). Even though the training step could be slow, but as soon as the classifier is trained, the classification is fast. For 1$ recognition, it does not have an explicit training step, instead, an alignment is required for every input, which is even more time-consuming.

Invariant Feature: The translation/scaling/rotation invariant feature is the best feature that people are looking for. 1$ recognizer, as I said, is not a feature-based method. However, the alignment step is applied to eliminate the effects of these transformations. But another problem rises, since for sketch, sometimes, these transformations are also an important feature for a certain meaning, for example, in Rubines’ paper, 4th feature, which is the angle between the bounding box diagonal and the bottom horizontal. This feature actually captured the rotation information, however, this information is lost in 1$ recognizer.