2010年9月2日星期四

Reading 1# : Gesture Recognition

Comments:
      Marty Field
Summary: 
Gesture recognition can relate to sketch recognition, only when sketches are drawn in the same manner every time. Then the paper surveys three common technologies on gesture recognition, Rubine, Long, Wobbrock 1$. It introduces the definition of features, why to use those features, figures of how to calculate features, how to build classifier, and the accuracy rate respectively. Also it makes some comparison among these three methods. Long introduced 11 new features, but didn't get better result. Wobbrock got better result by his $1 recognizer, but the time cost in recognition process was larger. Besides, it need more templates if sketches could be drawn in different ways.

Discussion:
In general, it is a very good survey to help beginners understand why and how gesture recognition is able to be used in sketch recognition.

About features: For some orientation-independent sketch recognition, orientation-independent features should be very useful, such as Hu invariant moments, which are robust to scale, rotation, translation and so on. I dont know whether they are applicable to gesture recognition, but I think they should be.

About time stamp: I think time stamp is very important in sketch recognition. For example, it is very useful to determine stroke orders. And all these three methods in the paper ignore time stamp to some extent.(not totally ignore) They simplify a 3-D problem to a 2-D problem, so that it will result in some inaccuracy in recognition. I think we may try to solve problems in 3-D coordinate. Maybe it is time-consuming,  or maybe problems are able to solve in 2-D coordinate.

Questions
1
The advantage of deleting the first or the second point is to get exact rotation angle without "divided by zero" error.

The disadvantage of deleting the first or the second point is to affect the calculation of the maximum speed. If the first point is deleted, the speed from the previous point of the first to the first will change to the speed from the previous point of the first to the second one. The time lasts longer, so the speed becomes smaller. So is the situation with deleting the second point. However, if the first point and the second point have the same time stamp, there will be no affection.
2.
The advantage of removing the first point or the second one is that we can still get the approximately max speed, while the disadvantage of that is to affect some features relating to rotation angle (such as the sum of rotation angle ).

The advantage of altering the time stamp is to get the approximate speed, while the disadvantage is to produce huge speed which is from the first point to the second point.
I think the best method maybe set a threshold and make Dtp = t(p)-t(p-1)+dtp. dtp is a very small time interval. If the max speed is more than the threshold, it should be caused by the same time stamp.

3 7-f 8-c 6-a 5-b 2-d 1-g 3-h 4-e

4 I think Long used the spirit of "density" to calculate kinds of density. Density here is in the wide definition. It can build kinds of distribution of the sketch in its area, in order to get more exact description of the sketch.

3 条评论:

  1. Hi Jianjie. I think Rubine makes good use of the timestamp to determine the maximum speed of the stroke and the total time it takes to complete the stroke (features 12 and 13). Long chose not to use the features related to time, but it didn’t hurt his results. What are some additional ways that you think these methods could have used the timestamp?

    回复删除
  2. I mean the dimension "t" should be used as "x" "y". I think it will increase the accuracy. Long just used 't' to calculate speed.

    回复删除
  3. What do you mena by dimension t as x and y? May be you are refering the 3D coordinates or what?

    Still I guess 2D sketch is a good starting point. I'm not sure how a 3D sketch can be convert to a 2D sketch!!!!

    回复删除