Reading #6

Comments:
Yue Li

Summary:


In this paper the author introduced a gesture recognizer - Protractor, which is an extension of 1$ recognizer, so they share a lot in common. So, I would like to use the similar four steps in 1$ recognizer to explain the Protector.

1. Points Resampling: N points with equal distance will be sampled.

2. Rotation: User can choose orientation-invariant or -sensitive. If -invariant is chosen, a similar rotation as in 1$ is done. If -sensitive is chosen, Protractor aligns the indicative orientation of a gesture with the one of eight base orientations that requires the least rotation. And this eight orientations are considered the major gesture orientations

3. Scaling and Translation: all the points will be translated to make the cancroids (0,0); not scaling is done, however, based on the way of doing Angular Distances, Protractor is inherently scale-invariant.

4. Calculating Optimal Angular Distances: Based on the vector representation obtained from previous steps, Protractor uses inverse cosine distances between their vectors as the similarity score. The template with the maximum similarity score will be the best match.

At last, the author gives many example, indicating the Protractor is superior to 1$ in terms of both accuracy and speed.



Discussion:

Protractor, as a template-based recognizer should share the similar property as 1$ recognizer. No training step is involved, and training samples are stored as templates. However, before classification, a preprocessing is needed, and the unknown gesture is compared against all the templates. This is time consuming, and will finally turns out to be slower than feature-based algorithm. For in this paper, the author did not make a comparison with feature-based, however, these properties should be similar.


As a comparison with 1$ recognizer:

Similar parts:

1. Points Resampling: They are doing the same way.

2. Translation: Still the same.

3. Scaling: Even though Protractor does not have rescaling step, according to the way of calculating the Angular Distance, we can assume a scale-invariant here.

4. Rotation-invariant: If this option is chosen, these two are both rotation-invariant.


Different parts;

1. Rotation-sensitive: Even though it is called rotation sensitive, we still need to eliminate the effect of rotation noise, instead of doing a full rotation, Protractor does a rotation to align to the eight orientations, for me this is kind of way, Rather than to totally align the rotation or to ignore the rotation noise, to align to eight orientations is kind like in-between.

2. Calculating Optimal Angular Distances: This part is the main contribution of the paper, the author indicates that the close-form solution for the calculation of vector based similar measurement is the key reasons why Protractor outperforms 1$ recognizer. This close-form solution has saved a lot of computation time rather than the iterative approach of finding rotation as in 1$ recognizer. And the results also indicates a better accuracy.