COMMENT
1.
2.
SUMMARY
The $3 recognizer is a simple, robust 3D gesture recognition method which relies on simple trogonometric and geometric calculations and requires less training samples. Authors feel that increasing number of mobile devices are equipped with 3D accelerometers. This 3D accelorometer data can be used in recognizing gestures.
The gesture data is first resampled so that points are redistributed to be at equal distances. To neutralize rotaional errors gestures, the gesture trace is rotated once along the gestures indicative angle (angle between the first points and the centroid). Lastly input gesture is scaled to fit a normalize cube. The input gesture is compared with all the gestures in the training data set. A scored list of candidate gestures is produced based on the mean square error (MSE). To reduce the occurence of false postitives the authors use a heuristic.
DISCUSSION
An interesting method due to simplicity and ease of implementation. However as mentioned by the authors in another paper describing the algorithm the method cannot recognize gestures in continious motion stream. There has to be an explicit start and end to the gesture.
Tuesday, March 30, 2010
Office activity recognition using hand posture cues
COMMENT
1.
2.
SUMMARY
This paper talks about identifying activities thorugh hand posture recognition. Paper's goal is to find out whether hand posture is a useful indicator for recognizing activities. It also tries to find out whether the hand postures can be generalized independent of the users.
Authors use a single right handed cyberglove with 22 censors reading the abduction and flexion angles of different joints in the hand. The paper uses a simple 1-nearest neighbor method to test hand posture recognition. The data for hand postures involved in activities was collected from 8 users while they were naturally performing activities around an office desk for e.g. Dialing a telephone, Using a mouse, reading a piece of paper etc. To test whether hand postures could be associated with correct activities independent of which user the recognizer was trained on, the authors performed a leave one out cross validation across all 8 users. The average accuracy in this case was 62.5%. For user dependent training and testing the accuracy obtained was 94.2%. This lead the authors to conclude that hand posture can indeed be used as a cue to perform activity based recognition. However, a user dependent trained system performs much better. A user independent system performs poorly due to high variation in how users carry out tasks.
DISCUSSION
The paper provides a good insight into how hand postures can inform activity recognition. It would be interesting to see if the method performs better when combined with vision techniques. For example its seems that the ambiguity in recognizing 'picking up a pape'r and 'picking up earphones' could be resolved if the system could see the 'object'..
2.
SUMMARY
This paper talks about identifying activities thorugh hand posture recognition. Paper's goal is to find out whether hand posture is a useful indicator for recognizing activities. It also tries to find out whether the hand postures can be generalized independent of the users.
Authors use a single right handed cyberglove with 22 censors reading the abduction and flexion angles of different joints in the hand. The paper uses a simple 1-nearest neighbor method to test hand posture recognition. The data for hand postures involved in activities was collected from 8 users while they were naturally performing activities around an office desk for e.g. Dialing a telephone, Using a mouse, reading a piece of paper etc. To test whether hand postures could be associated with correct activities independent of which user the recognizer was trained on, the authors performed a leave one out cross validation across all 8 users. The average accuracy in this case was 62.5%. For user dependent training and testing the accuracy obtained was 94.2%. This lead the authors to conclude that hand posture can indeed be used as a cue to perform activity based recognition. However, a user dependent trained system performs much better. A user independent system performs poorly due to high variation in how users carry out tasks.
DISCUSSION
The paper provides a good insight into how hand postures can inform activity recognition. It would be interesting to see if the method performs better when combined with vision techniques. For example its seems that the ambiguity in recognizing 'picking up a pape'r and 'picking up earphones' could be resolved if the system could see the 'object'..
Subscribe to:
Posts (Atom)
