From Pixels to Sentiment: Fine-tuning CNNs for Visual Sentiment Prediction

Visual media have become a crucial part of our social lives. The throughput of generated multimedia content, together with its richness for conveying sentiments and feelings, highlights the need of automated visual sentiment analysis tools. We explore how Convolutional Neural Networks (CNNs), a computational learning paradigm that has shown outstanding performance in several vision tasks, can be applied to the task of visual sentiment prediction by fine-tuning a state-of-the-art CNN. We analyze its architecture, studying several performance boosting techniques, which led to a network tuned to achieve a 6.1 % absolute accuracy improvement over the previous state-of-the-art on a dataset of images from a popular social media platform. Finally, we present visualizations of local patterns that the network associates to each image's sentiment.