Edited by author 11-02-2006 08:35 AM
(a)
Addiotnal Feature: use Complement Colors:
---Edge detection on Hue domain
---peak distance on hue distogram

(b)
edge distribution: spatial info. (no color info)
the others : global feature (consider colors but no spatial info.)
--> Combine them together.
--> Should apply spatial descriptor to various color space.

(c) Any idea finding professional quality video shots?

Coming up with low-level features to describe high-level concepts (such as simplicity and realism) seems to be doomed from the start. However, I really admire the authors for taking on such a difficult problem. I think for some of the more technique-based characteristics that define good photos there is real hope to be able to extract meaningful image features.

The features they use are actually fairly high level and thought out. I wonder if this could work in a boosting framework where you throw in thousands of really "dumb" features and let it combine them in a meaningful way, such that it discovers the high level concepts they are trying to model by itself.

I really liked how they came up with these features for something which at first seems pretty subjective. It'd be great to see an application of this in image retrieval.

Is there any information available on which quality features the snapshots from the tested dataset did worst on? since they used the bottom 10%, it seems like blur would be pretty common, making the other features not as relevant for that 10%. It'd also be interesting to see what kind of images accounted for the 24% error rate.

Interesting paper.. I'm curious how the artists' (photographers) react to this! I wonder if could be extended to look at qualities of DVDs. perhaps especially DVDs that have been produced from old B&W movies! =)

Adding image metadata as input to the classifier should help quite a bit.

Just knowing the model of the camera used to take the photo would probably help pick out "professional" photos more accurately than many of the perceptual criteria. It's almost hard to take a bad photo with some of the really expensive cameras, whereas it requires skill to take a professional quality photo with a cheap point and shoot digicam.

The authors seem to have engineered a pretty useful bag of tricks to tackle this difficult problem. I was surprised how much top-down design there was, and how little data-driven learning. I wonder how a state-of-the-art black box learner could do (either given these engineered features, given only more standard and basic features(which often contain the info in the engineered features indirectly), or given features somewhere in between).

Like some of the other comments, I would like to see if they could have learned the important features from the data instead of manufacturing them based on loose characteristic descriptions. They pride themselves on matching the performance of the previous work with less features, but they clearly spent a lot of time developing these more complicated featues specific to the application. Nonetheless, it is an interesting attempt at such a vaguely defined, but potentially useful classification problem.

Seems to me that photo quality is not a "high quality" vs "low quality" thing. It would make more sense to give an image a score and do some kind of regression from the data to give an incoming image a score from 0 to 1, where 1 is the highest quality. But I don't think I know how you would go about doing that.

Edited by author 11-02-2006 04:46 PM
I wonder if this works well when the collection of pictures is small, for example for the personal pictures, since the difference between "high quality" and "low quality" is less obvious. Why Do they use of RGB (and for the hue HSV)and not other colorspaces? any clue?