Today morning, the professor and me talked about the project we will work on, 3D-Daisy descriptor for action recognition. He has provided me a paper, 3D SIFT local descriptor, and hope me to extend the 2D Daisy descriptor into 3D, and works like 3D SIFT. At the beginning, I think the core of 3D-SIFT is very simple, even naive. Choosing some points in frames from a video stream and descript the information from those points. Then, by using k-means clustering algorithm, define a classifier to analyze the descriptors from points and classify them.
However, today, I think the problem is not as simple as I originally supposed. If I want to be accepted by a conference or committee, I must show some new ideas in my works or some amazing result... I do not think the 3D daisy, if as I supposed what it should be, has anything new or break through...what should I do?
Firstly, Mr. Guo wants to design a "cutter". This cutter can reduce or even cut the weight of the region we not interested. I think this method is good, but hard to implement. Our computer cannot know where is the region we interested. If the computer cuts any region where may provide information to help the classifying, it means the purpose of the cutter is fail to get.
Secondly, we believe that, for walking, the movement of legs and hands is important, but we cannot add any additional knowledge of the structure of the body...If we do so, our program will fall into the domain of tracing and geometry, where is a totally different subject to research!