PieAPP: Perceptual Image-Error Assessment through Pairwise Preference.

🌏 Source

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

- 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.

**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 .

- papers
- | Paper Title | Publication | Source Code | | perceptual-similarity | CVPR 2018 | richzhang/PerceptualSimilarity | | pieapp | CVPR 2018 | prashnani/PerceptualImageError |