🍌 Spencer's Wiki

# PieAPP

PieAPP: Perceptual Image-Error Assessment through Pairwise Preference.

🌏 Source

Available at https://arxiv.org/abs/1806.02067, source code available at: prashnani/PerceptualImageError.

# Brief introduction

  • The paper proposes a new, large-scale dataset labeled with the probability that humans will prefer one image over another.
  • Then the paper trains a deep-learning model using a novel, pairwise-learning framework to predict the preference of one distorted image over the other.

The new metric: PieAPP, correlates well with human opinion, and performs almost 3 times better than existing metrics.

# Main contributions

  • Dataset: The paper doesn't explicitly convert the human preference into a quality score. Instead, they simply label the pairs by the percentage of people who preferred image over .

  • Pairwise-learning framework: The framework trains an error-estimation function using the probability labels in the dataset.

    ---------------------
| Distorted Image A | \
---------------------  \
                        \
---------------------    \  | Error estimation function #1 |       | Perceptual Error Score: A |
| Reference Image R | -->-- |--------- IDENTICAL ----------| -->-- |---------------------------|
---------------------    /  | Error estimation function #2 |       | Perceptual Error Score: B |
                        /
---------------------  /
| Distorted Image B | /
---------------------

    Then the errors of and are used to compute the predicted probability of preference for the image pair.

  • Once the PieAPP (described above) is trained using the pairwise probabilities, we can use the learned error-estimation function on a single image and its reference to compute the perceptual error of with respect to .

# Method overview

# Experiments

# Referred in