Review for Paper: Priors for Large Photo Collections and What they Reveal about Cameras

Reviewed by: Changyin Zhou, cz2166


Part 1
======

Overall Rating:
	__Definitely Accept

Explanation:

(Please summarize the paper and describe its key contributions in your
 own words. Please stress both the positive and negative aspects of the
 paper. In particular, please explicitly describe what you feel is the
 best thing about this paper, and what is the worst.)


In this paper, the authors studied the aggregate statistics of large photo collections. While most previous work on image statistics focus on the statistics of individual photos (e.g. 1/f law and sparsity of image derivatives), this work demonstrates another dimension to look at the image statistics. Although these aggregate statistics, more or less, is an obvious fact to some researchers, it is still nice to see a deep and solid invistigation.  

The authors also suggested to use this statistics to estimate response functions and correct vignetting for specific camera models, and to detect bad pixels for specific camera instances. For the simplicity of the algorithm and the availability of large photo collections on the Internet, their method based on aggregate statistics could be a wise choice.

The authors have done a great job in explaining the intuition behind the idea, performing experimental evaluation, and exploiting possible applications of the aggregate statistics. However, they did not explicitly formulate the aggregate statistics and did not conduct some necessary statistical analysis. For example, how will the correlation among camera models, camera settings, scenes, and photographers affect the aggregate statistics? It seems that the authors have implicitly assumed the independence among these factors throughout the whole discussion. But note that these correlations could be stronge in many situations. Overlooking these correlations could lead to severely biased estimations.



Part 2
======

Confidence:
	__Confident


Explanation:

(The default selection is "Confident". Use the other options to stress
 that you are absolutely sure about your conclusions (you are an expert
 in the respective area) or that you feel some doubt. If you have
 serious doubts about your ability to assess the paper, please complete
 the review as best you can, but make a note of this in the private
 comments to the committee below.)




Part 3
======

Novelty
	__Original


Explanation

("Very Original" papers open new directions and often become seminal
papers. "Has Been Done Before" must be accompanied by relevant
references.)

The idea of aggregate statistics itself is not so original. However, the authors investigated in this idea and showed several novel observations and novel applications.


Part 4
======

Importance
	__Interesting to a Subarea

Explanation

(Papers "Of broad interest" should be of interest to most of the
community, e.g., for the technical quality of the work, or for a
surprising or particularly impressive result. Papers "Interesting to a
Subarea" do not have to address every attendee, but should have an
impact in a certain area. Papers "Interesting Only to a Small Number of
Attendees" should only be considered for publication if they are very
strong in other aspects.)

It could be interesting to some people who do image enhencement and who do research in large photo collections.





Part 5
======

Reference to Prior Work
	__Excellent

Explanation

("Misses Key Related Work" strongly suggests reject. This option should
only be selected if the missing work is well known in the community and
commonly cited. Otherwise you should give the authors the benefit of the
doubt by selecting "Some References Missing" and by providing pointers
to the missing references. The authors can easily add missing
references, and will be expected to do so.)





Part 6
======

Clarity of Presentation
	__Reads Very Well

Explanation

(Everything else being equal, clearly written papers are preferred.
However, a paper with strong original results should not be rejected
only because it is not well written. If you were able to evaluate the
preceding criteria and to explain the key contribution, the work is
definitely not "Unreadable". On the other hand, papers that are
completely unintelligible can be rejected for this reason alone.)




Part 7
======

Technical Correctness
	__Definitely Correct

Explanation

(The statement that a paper is "Definitely Correct" means that its
conclusions are supported by flawless arguments. Proofs are correct,
formulas are correct, there are no hidden assumptions,
experiments are well designed and properly evaluated. The
important distinction to make in this section is between
errors that can be rectified relatively easily in the final
version, and errors that invalidate the entire approach.)







Part 8
======

Experimental Validation
	__Excellent Validation or N/A-a Theoretical Paper

Explanation

(Different papers need different levels of experimental validation.
Please take this into account while completing this section. A
theoretical paper may need no experiments. A paper presenting a new idea
might just need an experiment illustrating that there exists a situation
where the idea applies. A paper presenting a performance evaluation
paper may need extremely thorough experiments and their evaluation. It
is up to you to judge what is appropriate for each paper.)



Part 9
======

Additional Comments: