Citation:
Wobbrock, Jacob O., Andrew D. Wilson, and Yang Li. "Gestures without libraries, toolkits or training: a $1 recognizer for user interface prototypes."Proceedings of the 20th annual ACM symposium on User interface software and technology. ACM, 2007.
Publication Link
Summary:
This paper presents an entirely new view on gesture recognition, and openly questions all the universally accepted methods up till now, pointing out that they are slow, complicated, inaccessible and domain restricted. A new technique is presented, and is evaluated against its counterparts to prove that it does equally well if not better. The paper also claims that the entire algorithm can be implemented in 100 lines of codes, making it very easy for beginners to pick up and use in their work.
Discussion:
The aim of this paper to present a simple yet effective gesture recognition algorithm that does not delve into pattern recognition algorithms.
The 4-step algorithm used is:
1. Resample the point path (generally 64 equidistant points)
2. Rotating once based on indicative angle (angle between centroid and first point)
3. Scale (non-uniformly onto a square) and Translate (to origin)
4. Find the optimal angle for the best score (Finds average distance between corresponding points in candidate and the template being checked against)
To maintain rotation invariance, it was found that for similar gestures, the indicate angle was almost accurate, and a simple hill climbing algorithm from there would take us to the global minima (local minima effect is low). However, this ended up in too many rotation steps in the case of dissimilar gestures. GSS(Golden Section Search) using the Golden Ratio is used to amend this at the cost of some performance for similar gestures.
Limitations of $1 Recognizer:
- 1D figures
- Overdependent on non-uniform scaling
- Assumes rotation and translation invariance, and hence cannot capture specifics in these aspects
- Does not take time into account
The $1 recognizer was compared against the DTW, which was inefficient and scaled poorly, as well as Rubine's method which proved to require too many examples to show some acceptable accuracy. The $1 recognizer on the other hand, was much faster and achieved accuracy with minimal training requirement.
The factors considered for experimentation are:
1. No. of templates for training
2. Gesture speed
3. Separation boundaries between competing templates
Wobbrock, Jacob O., Andrew D. Wilson, and Yang Li. "Gestures without libraries, toolkits or training: a $1 recognizer for user interface prototypes."Proceedings of the 20th annual ACM symposium on User interface software and technology. ACM, 2007.
Publication Link
Summary:
This paper presents an entirely new view on gesture recognition, and openly questions all the universally accepted methods up till now, pointing out that they are slow, complicated, inaccessible and domain restricted. A new technique is presented, and is evaluated against its counterparts to prove that it does equally well if not better. The paper also claims that the entire algorithm can be implemented in 100 lines of codes, making it very easy for beginners to pick up and use in their work.
Discussion:
The aim of this paper to present a simple yet effective gesture recognition algorithm that does not delve into pattern recognition algorithms.
The 4-step algorithm used is:
1. Resample the point path (generally 64 equidistant points)
2. Rotating once based on indicative angle (angle between centroid and first point)
3. Scale (non-uniformly onto a square) and Translate (to origin)
4. Find the optimal angle for the best score (Finds average distance between corresponding points in candidate and the template being checked against)
To maintain rotation invariance, it was found that for similar gestures, the indicate angle was almost accurate, and a simple hill climbing algorithm from there would take us to the global minima (local minima effect is low). However, this ended up in too many rotation steps in the case of dissimilar gestures. GSS(Golden Section Search) using the Golden Ratio is used to amend this at the cost of some performance for similar gestures.
Limitations of $1 Recognizer:
- 1D figures
- Overdependent on non-uniform scaling
- Assumes rotation and translation invariance, and hence cannot capture specifics in these aspects
- Does not take time into account
The $1 recognizer was compared against the DTW, which was inefficient and scaled poorly, as well as Rubine's method which proved to require too many examples to show some acceptable accuracy. The $1 recognizer on the other hand, was much faster and achieved accuracy with minimal training requirement.
The factors considered for experimentation are:
1. No. of templates for training
2. Gesture speed
3. Separation boundaries between competing templates
No comments:
Post a Comment