Comments on:
Wenzhe Li
Summary:
The problem of similarity always confuses people and computer as well. However, it is difficult for interface designers to solve this problem by themselves. This paper provides a tool named “quill”, which will automatically give advices to designers when it thinks is necessary.
Based on the similarity metrics that the author built from three experiments, quill predicts when the similarity problem would occur and offers unsolicited warnings and advices to designers.
The author found interface challenges in three areas: when to advice, how much to advice and what to advice. During the implementation stage, the author chose different strategies for user-initiated and system-initiated analyses. And because of the similarity metric, quill inclines to overestimate similarity when dealing with letter related gestures.
Discussion:
The author shows almost all the matters of developing software: what is its aim, the challenges of interface, implementation and underlying model (not perfect in this case). I realize that there is many things to consider when developing a software rather than just finishing its core algorithm and then being happy and satisfied. The author concerned a lot about details, which may not be very interesting and exciting but still need thorough consideration.
2010年9月8日星期三
2010年9月7日星期二
Reading #1 Gesture Recognition
Reading #1 Gesture Recognition
COMMENTS ON:
youyou wang
SUMMARY:
This paper is an introduction to gesture recognition and its three representative methods. In order to recognize gestures, the gesture recognition asks their path to remain the same, which is a main limitation.
Among the shown tree method, Rubine’s method and Long’s are based on features of the path. Rubine lists 13 features and the last two describe the feature of the max speed and the time used. While Long do not think these two feature is insightful and replace them with another 11 features. And both of these two gesture recognition need training set for their linear classifier to wok well.
Wobbrock’s recognizer is to transform the original gesture (in order to better comparing the template) and then match it to specific template (by calculating their “distance”).
DISCUSSION:
The computer needs to know enough information of the gesture in order to recognize it. This information can be achieved through two different ways: The computer learns it from its experiences or directly given by the user. Rubine and Long follow the former one, defining the feature that they thought adequate to recognize a path and then throwing input to train the recognizer. While Wobbrock chose to give the computer a database of some candidate templates, transform the input and try to match it to the database.
So which way is a better way? And is accuracy the only criteria? Wobbrock’s method tends to perform better but cost more time and show limitation when facing the variation of the same gesture. So, “to teach people to learn is more important than to teach people to know” also applies on gesture recognition?
COMMENTS ON:
youyou wang
SUMMARY:
This paper is an introduction to gesture recognition and its three representative methods. In order to recognize gestures, the gesture recognition asks their path to remain the same, which is a main limitation.
Among the shown tree method, Rubine’s method and Long’s are based on features of the path. Rubine lists 13 features and the last two describe the feature of the max speed and the time used. While Long do not think these two feature is insightful and replace them with another 11 features. And both of these two gesture recognition need training set for their linear classifier to wok well.
Wobbrock’s recognizer is to transform the original gesture (in order to better comparing the template) and then match it to specific template (by calculating their “distance”).
DISCUSSION:
The computer needs to know enough information of the gesture in order to recognize it. This information can be achieved through two different ways: The computer learns it from its experiences or directly given by the user. Rubine and Long follow the former one, defining the feature that they thought adequate to recognize a path and then throwing input to train the recognizer. While Wobbrock chose to give the computer a database of some candidate templates, transform the input and try to match it to the database.
So which way is a better way? And is accuracy the only criteria? Wobbrock’s method tends to perform better but cost more time and show limitation when facing the variation of the same gesture. So, “to teach people to learn is more important than to teach people to know” also applies on gesture recognition?
订阅：
博文 (Atom)