LAPIS: A novel dataset for personalized image aesthetic assessment

TL;DR

LAPIS presents the first large-scale, artistically focused dataset for personalized image aesthetic assessment, addressing biases in prior IAA datasets by combining rich image attributes with extensive annotator personal attributes. The authors curate 11,723 WikiArt images with 31 image attributes and 47 features, rated by about 24 annotators per image, and collect demographic and interest data to enable PIAA modeling. They evaluate GIAA with a ResNet-50 baseline and implement two PIAA models (PIAA-MIR and PIAA-ICI), showing that personal and image attributes improve predictions, though cross-annotator generalization remains challenging. Ablation studies reveal the importance of art interest and style/genre information, while failure analyses highlight limitations in predicting for unseen users and certain genres, motivating future work on generalization and model design. Overall, LAPIS establishes a high-quality resource for advancing artistic PIAA and supports more nuanced personalization in aesthetic predictions for art-related applications.

Abstract

We present the Leuven Art Personalized Image Set (LAPIS), a novel dataset for personalized image aesthetic assessment (PIAA). It is the first dataset with images of artworks that is suitable for PIAA. LAPIS consists of 11,723 images and was meticulously curated in collaboration with art historians. Each image has an aesthetics score and a set of image attributes known to relate to aesthetic appreciation. Besides rich image attributes, LAPIS offers rich personal attributes of each annotator. We implemented two existing state-of-the-art PIAA models and assessed their performance on LAPIS. We assess the contribution of personal attributes and image attributes through ablation studies and find that performance deteriorates when certain personal and image attributes are removed. An analysis of failure cases reveals that both existing models make similar incorrect predictions, highlighting the need for improvements in artistic image aesthetic assessment. The LAPIS project page can be found at: https://github.com/Anne-SofieMaerten/LAPIS

Figures (19)

• Figure 1: Illustration of the image selection. Images on the left were excluded during the quality check. The top left image contains a watermark and the bottom left image is a sculpture. The images on the right are example images in LAPIS.
• Figure 2: Visualisation of the types of data in LAPIS. All images have metadata (title, artist) and image attributes. Images are rated by multiple users on their aesthetic appeal. For each user, we have a set of personal attributes.
• Figure 3: Violin plots of the data distribution per style. Violins are ordered from lowest median (top) to highest median aesthetic scores (bottom). The abstract and figurative styles are shown in different shades of red and blue respectively.
• Figure 4: Scatter plot of the mean aesthetic rating given by an annotator in function of their art interest score . The marginal distributions for both art interest and aesthetic scores are shown on the side. We found a correlation between art interest and aesthetic scores ($r=0.35, p < 0.01$).
• Figure 5: Histogram of the aesthetic scores averaged per image. Data corresponding to abstract and figurative works is shown in red and blue respectively. We observe a trend towards preferences for figurative works.