Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generalized Portrait Quality Assessment

Nicolas Chahine, Sira Ferradans, Javier Vazquez-Corral, Jean Ponce

TL;DR

This paper addresses portrait quality assessment (PQA) under diverse real-world scenes. It introduces FULL-HyperIQA (FHIQA), a semantics-aware BIQA model that uses a scene-prediction vector to perform scene-conditioned rescaling of patch-level quality with the final score $Q_f = ( \sum_{i=1}^{k} P_{s_i} (a^{s}_{i} Q_p + b^{s}_{i}) ) / ( \sum_{j=1}^{k} P_{s_j} )$. It extends prior HyperIQA with a full-scene prediction for handling unseen content and shows competitive or superior performance on PIQ23 across multiple attributes, especially overall portrait quality. This demonstrates the value of semantic understanding for robust, generalizable IQA in smartphone photography and proposes a path toward content-specific evaluation.

Abstract

Automated and robust portrait quality assessment (PQA) is of paramount importance in high-impact applications such as smartphone photography. This paper presents FHIQA, a learning-based approach to PQA that introduces a simple but effective quality score rescaling method based on image semantics, to enhance the precision of fine-grained image quality metrics while ensuring robust generalization to various scene settings beyond the training dataset. The proposed approach is validated by extensive experiments on the PIQ23 benchmark and comparisons with the current state of the art. The source code of FHIQA will be made publicly available on the PIQ23 GitHub repository at https://github.com/DXOMARK-Research/PIQ2023.

Generalized Portrait Quality Assessment

TL;DR

This paper addresses portrait quality assessment (PQA) under diverse real-world scenes. It introduces FULL-HyperIQA (FHIQA), a semantics-aware BIQA model that uses a scene-prediction vector to perform scene-conditioned rescaling of patch-level quality with the final score . It extends prior HyperIQA with a full-scene prediction for handling unseen content and shows competitive or superior performance on PIQ23 across multiple attributes, especially overall portrait quality. This demonstrates the value of semantic understanding for robust, generalizable IQA in smartphone photography and proposes a path toward content-specific evaluation.

Abstract

Automated and robust portrait quality assessment (PQA) is of paramount importance in high-impact applications such as smartphone photography. This paper presents FHIQA, a learning-based approach to PQA that introduces a simple but effective quality score rescaling method based on image semantics, to enhance the precision of fine-grained image quality metrics while ensuring robust generalization to various scene settings beyond the training dataset. The proposed approach is validated by extensive experiments on the PIQ23 benchmark and comparisons with the current state of the art. The source code of FHIQA will be made publicly available on the PIQ23 GitHub repository at https://github.com/DXOMARK-Research/PIQ2023.
Paper Structure (12 sections, 1 equation, 4 figures, 1 table)

This paper contains 12 sections, 1 equation, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Scene-specific rescaling
  4. Understanding new content
  5. Dataset
  6. Experiments
  7. Baselines methods
  8. Training strategy
  9. Metrics
  10. Results
  11. Understanding new content
  12. Conclusion

Figures (4)

  • Figure 1: Diagram of FULL-HyperIQA (FHIQA). The figure illustrates how FHIQA processes input images, extracts semantic information, and adapts the quality prediction based on scene-specific evaluations.
  • Figure 2: Examples from the new scene split for PIQ23. The test set incorporates various framing settings, backgrounds, subject characteristics, and weather conditions that are significantly distinct from the training set.
  • Figure 3: Comparative analysis of IQA models based on the averaged correlation metrics distribution across all scenes and for the three attributes of PIQ23.
  • Figure 4: Histograms showing the classification distribution across training scenes for various unseen testing conditions. The same testing scene can be projected to multiple training scenes with similar features, showcasing the necessity to consider multiple scenes for inference on new conditions.