Review (HW5):
All of you did a good job, coming to similar results (with different conclusions, though), i therefore will not give separate comments.
Following your suggestions, i would have to decide to rank the paper as 'major revision', which is reasonable in my opinion, too.
Some comments:
- Nearly all of you mentioned the simplicity of the approach as a positive aspect - i agree.
- All of you mentioned the insufficient experiments - i wholeheartedly agree.
- The readability ranked from 'hard to understand' to well written, but readability is very subjective anyway. I personally find it very readable, even a bit 'over detailed' (the rotation formula...)
- One comment i liked: 'i tend to reject papers with more than five parameters, if there is no analysis'. That's exactly the point of this paper: it shows a simple approach, brute force implemented, yet no analysis. What's happening here is the following: given a shape, we lose information by feature extraction (here: shape context), and, as we learned in class, in a second step, the feature comparison (question: which metric to use?). This paper adds a step in between: in transforms to a different feature space (Fourier space), which by itself does not lose any information, but is a different representation which calls for analysis of the characteristics of the new space: shapes that were close in the original, shape context feature space might not be close any longer, and vice versa. Hence it is important to either analyze the properties of the conversion (which might be hard), or at least to give sufficient examples of classes of shapes, that might give a hint about what's going on in the FFS space.
- Of course, looking at the spectrum only, loses further information, that's exactly where we gain the rotational invariance.