# Understanding Image Virality

**Authors:** Arturo Deza, Devi Parikh  
**arXiv:** [1503.02318](https://arxiv.org/abs/1503.02318)  
**Date:** 2015-03-08  
**Summary page:** https://www.summarizepaper.com/en/paper/1503.02318v3/

## Key points

- The paper explores the phenomenon of virality in online content, specifically focusing on viral images from a computer vision perspective.
- Three new image datasets sourced from Reddit are introduced.
- A virality score based on Reddit metadata is developed.
- Classifiers using advanced image features are trained to predict the virality of individual images, relative virality between pairs of images, and dominant topic of a viral image.
- Computers struggle when using low-level features alone, emphasizing the importance of high-level information for accurate prediction.
- Relative attributes are encoded into models to incorporate semantic information.
- Five key visual attributes that strongly correlate with virality are identified.
- An attribute-based characterization system achieves 68.10% accuracy in predicting relative virality, surpassing human performance at 60.12%.
- Human predictions of image virality vary in different contexts such as neighboring images, recently viewed images, and image titles or captions that may influence perception of viral potential.
- The authors plan to make their datasets and annotations publicly available for further research and exploration.

## Summary

The paper "Understanding Image Virality" by Arturo Deza and Devi Parikh explores the phenomenon of virality in online content, specifically focusing on viral images from a computer vision perspective. The authors introduce three new image datasets sourced from Reddit and develop a virality score based on Reddit metadata. They train classifiers using advanced image features to predict the virality of individual images, the relative virality between pairs of images, and the dominant topic of a viral image. In their study, the authors compare the performance of machine classifiers to human performance on these tasks. They discover that computers struggle when using low-level features alone and emphasize the importance of high-level information for accurately predicting virality. To incorporate semantic information, they encode relative attributes into their models. Through their research, the authors identify five key visual attributes that strongly correlate with virality. They create an attribute-based characterization system for images that achieves 68.10% accuracy in predicting relative virality (using SVM+Deep Relative Attributes), surpassing human performance at 60.12%. Additionally, the paper investigates how human predictions of image virality vary in different contexts such as neighboring images, recently viewed images, and image titles or captions that may influence perception of viral potential. Overall, this work represents an initial step towards comprehending the complex yet significant phenomenon of image virality. The authors plan to make their datasets and annotations publicly available for further research and exploration in this field.