Image Aesthetic Assessment Based on Pairwise Comparison A
Image Aesthetic Assessment Based on Pairwise Comparison A Unified Approach to Score Regression, Binary Classification, and Personalization ICCV 2019 (Poster @ Applications. Medical, & Robotics) Jun-Tae Lee and Chang-Su Kim Media Communications Lab, Korea University Presenter: Dingquan Li
Outline • Problem • Background & Motivation • Method • Experiments • Summary
Outline • Problem • Background & Motivation • Method • Experiments • Summary
Problem
Outline • Problem • Background & Motivation • Method • Experiments • Summary
Background & Motivation • Most works focus on binary aesthetic classification while few works focus on aesthetic score regression and personalized aesthetics. • The latter ones are more challenging than the former one. • The latter ones are needed in some applications, such as image recommendation, image enhancement, and personal album curation. Can these three tasks be unified?
Outline • Problem • Background & Motivation • Method • Experiments • Summary
Method • Aesthetic Comparator for Pairwise Comparison (PC) • Aesthetic Score Regression Based on PC Matrix • Binary Aesthetic Classification by Comparing with Middles • Personalized Image Aesthetics by Extending the Generic One
Aesthetic Comparator
Feature Extractors
Ternary Classifier: Ground Truth Classes Lloyd, Stuart. "Least squares quantization in PCM. " IEEE Transactions on Information Theory 28. 2 (1982): 129 -137.
Ternary Classifier: Loss Function
Method • Aesthetic Comparator for Pairwise Comparison (PC) • Aesthetic Score Regression Based on PC Matrix • Binary Aesthetic Classification by Comparing with Middles • Personalized Image Aesthetics by Extending the Generic One
Constructing PC Matrix
Constructing PC Matrix
PC Matrix Example
Priority Vector of Aesthetic Scores Saaty, Thomas L. "A scaling method for priorities in hierarchical structures. " Journal of Mathematical Psychology 15. 3 (1977): 234 -281.
Aesthetic Score Vector Priority vector Score vector
Estimate Scale Factor
Aesthetic Score Regression
Reference Image Selection •
Method • Aesthetic Comparator for Pairwise Comparison (PC) • Aesthetic Score Regression Based on PC Matrix • Binary Aesthetic Classification by Comparing with Middles • Personalized Image Aesthetics by Extending the Generic One
Method • Aesthetic Comparator for Pairwise Comparison (PC) • Aesthetic Score Regression Based on PC Matrix • Binary Aesthetic Classification by Comparing with Middles • Personalized Image Aesthetics by Extending the Generic One
Extend the Generic One for Personalization
PC Matric Example
Outline • Problem • Background & Motivation • Method • Experiments • Summary
Experiments • Datasets • Aesthetic Score Regression • Binary Aesthetic Classification • Personalized Image Aesthetics
Datasets • AVA (CVPR 2012): 235599 images for training (choosing 2000 for validation) and 19930 images for testing. Ratings are from 110. • AADB (ECCV 2016): 8500/1000 images for training/validation/testing. Score ranges are [0, 1]. • FLICKER-AES (ICCV 2017 ): 35263 images for training and 4737 images for testing. Score ranges are [0, 5]. • Note: there is no overlap between workers of training image and testing images (173+37=210). As for the test images, each worker rated about 137 images.
Experiments • Datasets • Aesthetic Score Regression • Binary Aesthetic Classification • Personalized Image Aesthetics
(Qualitative) Examples (ground-truth, regressed score)
Quantitative Results
Comparison Examples of the Ternary Classifier
Finer Quantization in Aesthetic Comparator ?
Experiments • Datasets • Aesthetic Score Regression • Binary Aesthetic Classification • Personalized Image Aesthetics
(Qualitative) Examples
Quantitative Results on AVA
Experiments • Datasets • Aesthetic Score Regression • Binary Aesthetic Classification • Personalized Image Aesthetics
(Qualitative) Examples (the worker’s annotated score, regressed generic score, regressed personalized score)
Quantitative Results on FLICKER-AES Spearman’s Coefficient
Outline • Problem • Background & Motivation • Method • Experiments • Summary
Summarize Its Contribution and Limit •
- Slides: 42