A agree with Paul in the observation that domain specific knowledge was used VERY heavily in this paper.

In particular, the assumption of symmetry in left/right limbs is very domain specific and honestly isn't always true. As an example, consider a hockey goalie who has different padding on each arm.

Also, It seems like some more accuracy could be gained from feedback from the pose to the segmentation as long as you're doing heavily domain specific tricks. There are many times in their sample images where the segmentation could be greatly enhanced by examining the recovered pose. As an example, in figure 10 on the second row, the missing portion of the torso and leg could be found by just having a min-width associated with the stickman's middle, and upper leg.

In any case, this was a fun paper, so I don't mean to poke at it too much.

I like their intuition of starting with the half-limbs as "islands of saliency." Any idea as to the time it takes to process an image? There seems to be quite a lot of combinatorial searches going on.

Base on this paper, we know the importance of choosing good examples.
(a) The photos are taken by professional photographer for news. So the target will be in-focus while the background being out-of-focus.
(b) The joint of arms can be easily determined by the length of sleeves.

The performance can be improved using multiple images of same person with different
(a) time
(b) viewpoint

WARNING: No question just a comment!

Very interesting as we read these various papers how each system uses varying level of domain specific knowledge to achieve its goal. This system definitely falls on the lots of domain specific knowledge side. I like this aspect of their approach as it seems to mimic some of the higher-level constraints that our visual system applies to understanding natural scenes.

Edited by author 11-09-2006 12:32 AM
--Questions--

1. In the caption of Figure 4: What do they mean to have "relative scale and symmetry in clothing"?

2. Under "Shading Cue"

 a. what does it mean for the gradient images to be "half-wave rectified"?

 b. Is Figure 5 displaying the segmented limbs in the training set (not the mean of all shading descriptors)? The wording was confusing & I'm not sure what shading descriptors look like!

Thank you!!!

--Comments--

1. I REALLY like how they used a sigmoid function to transform each cue into a prob.-like quantity for the normalizing of features before logsitic regression. I have never heard of this before! Is this a common thing to do in Computer Vision?

2. I would love to see how well Shaknarovich et el's fast pose estimation algorithm does with the baseball people image set using Locality Sensitive Hashing! That would be great to be fast AND find the baseball people accurately with all their funny poses!