Dataset: 3581 images from Photo'net Datta et al', 2006

1
Learning the Consensus on Visual
Quality for Next-Generation Image Management
Ritendra Datta, Jia Li, and James Z. Wang The
Pennsylvania State University, University Park,
PA, USA
PROBLEM SPECIFICATION
EXPERIMENTS
METHODOLOGY

• Dataset 3581 images from Photo.net Datta et
  al., 2006
• Aesthetics scoring scale 1-7
• We have averages scores and count nk of
  ratings for Ik
• Characteristics of Dataset
• Results

• Formally, define consensus on quality for image
  Ik as
• where qk,i is the ith user rating received
• Actually, estimated over a finite set of ratings
  ?
• Useful for the average user Our target
  audience
• May not be useful for the outlier user
  Personalized recommendations?
• Let us have D visual features
  that correlate with quality
• In the literature Datta et al., ECCV 2006
  56 features higher-order terms
• Brightness, Contrast, DOF, Saturation, Region
  Composition, etc.
• (A) Learn a Weighted Least Squares Linear
  Regressor
• Directly learn a mapping from features to
  consensus scores
• Weights related to trust associated with score
• Greater the no. of samples n, better estimate
  is of
• Formulation
• Convenient parameter estimation
• Here,
  and denotes pseudoinverse.

• Photo-sharing is getting popular
• Dedicated Websites
• Flickr, Photobucket, Photo.net, etc.
• Bundled with social networking sites
• Facebook, Orkut, MySpace, etc.
• Personal Photo Collections are growing
• Cheap digital cameras and storage
• Organizing them can be time-consuming
• Many methods for topic-based management
• Image classification
• Content-based image search/annotation
• Collaborative tagging architectures
• How about quality-based photo management?
• Can we distinguish these photos
• (Rated over 6 out of 7 on average by many
  Photo.net users)
• from these ones?

On an average, picks 20 high-quality (HIGH5.5)
photos with 82 precision, while SVM-based
method gets less than 50
On an average, eliminates 50 low-quality
(LOW4.5) photos with 65 precision, while
wrongly eliminating only 9 high-quality photos.
The SVM figures are 43 and 28 resp.
Selection/Elimination
Sort Descending
HIGH QUALITY
CONCLUDING REMARKS
Apply Regressor Predict N scores

• Encouraging results, moving toward real-world
  applicability.
• Proposed features in Datta et al. 2006 show
  more promise.
• Weighting training data by confidence improves
  performance.

Sort Ascending
Collection of N images
LOW QUALITY