I noticed that they didn't have any examples or real discussion (that I remember) of multi-object single background segmentation. Is that current work? Does it seem to anybody else like this algorithm couldn't be easily extended to do that?
(btw example would be two elephants on a savannah)

They claim that the coarse estimate provides robustness to texture and clutter, but they use for training and testing only images that contain at least one salient, *unoccluded" object that is fully contained in the image.

Can you explain the intuition of how Fourier coefficients represent a shape? Thank you!

I am curious as to how similar the probability distributions found from the training data (e.g figures 1& 2) from the Berkely Segmentation Database will be to other training sets of natural images.

I.e. do these distributions work only for this particular dataset or do we have to re-train for every dataset?

From the paper:
Run-time for
our current (Matlab) implementations is between 45 min.
and 1 hr. for the multi-scale method, and between 3 and
5 min. for the single-scale version. RC and EJ (both of
which are binary packages) take less than 1 min. to complete.

So their algorithm takes 45 min to run on 2 image? Or am I misunderstanding it? That seems very long.

-Nadav

-I would would like to see how the other algorithms performed on the downsampled images. They handpicked their data for it to give the most information at low resolution.
-What are the applications of a very slow segmentation algorithm that only works well on previously picked easy data?
-Are the given results typical of RC and EJ algorithm segmentations? I don't know much about these algorithms, but it seems like they must be able to perform better than what is shown in fig 4.

Edited by author 10-24-2006 08:08 AM
To Boris:
For other class of images, depending on
(a) the underlying boundary detector in [15] which can be retrained.
(b) if multiscale(MS) is beneficial.

Some my points.
(a) Why restrict the application to only natural images? (due to the MS property?)

(b) Taking top 20 candidate contours in testing hides the high FP rate problem. In data mining view, the author cheated by looking at testing data!.

(c) Fig.4 doesn't agree with Fig.5. MS=0.5(too high) while EJ=0.73(too low) and SS>EJ, huh? The examples selected in Fig.4 is extremely biased to demo the MS algorithm.

(d) The grouping algorithm which always keep top K candidate after branch has a fancy name in NLP called "beam search".

(e) What's the major benefit of using Fourier Transform, does it really save computation time? Can you briefly describe the time that FFT used.

Though the authors write "In
this paper we are interested in computing the boundaries of
a broad range of salient objects", I wonder how this method would perform if it was trained on one specific class of objects like human skin, etc.

Edited by author 10-24-2006 12:39 AM
I like the paper's contribution of a way to integrate information at multiple scales. However, while it is a good idea, whether or not the framework is of any use depends on solving extremely difficult problems. These problems are mentioned in the discussion section and include defining the correct goodness metric for contours and developing a shape model of natural scenes. This difficulty is highlighted by the fact that after so much fanfare about using information at higher resolution the authors downsample the training images in order to improve results. Just an observation, not much of a question here.

It seems as though this approach could suffer from early thresholding in generating its set of Tangent segments T, since the algorithm determines what to include in this set prior to trying to connect it with its neighbors.

Also, I was just curious what 'Markov assumptions that do not capture global structure' many other boundary detection algorithms used (mentioned in the introduction).