Cross-Media Learning for Image Sentiment Analysis in the Wild
Lucia Vadicamo, Fabio Carrara, Andrea Cimino, Stefano Cresci, Felice Dell'Orletta, Fabrizio Falchi, Maurizio Tesconi


Much progress has been made in the field of sentiment analysis in the past years. Researchers relied on textual data for this task, while only recently they have started investigating approaches to predict sentiments from multimedia content. With the increasing amount of data shared on social media, there is also a rapidly growing interest in approaches that work ``in the wild'', i.e. that are able to deal with uncontrolled conditions. In this work, we faced the challenge of training a visual sentiment classifier starting from a large set of user-generated and unlabeled contents. In particular, we collected more than 3 million tweets containing both text and images, and we leveraged on the sentiment polarity of the textual contents to train a visual sentiment classifier. To the best of our knowledge, this is the first time that a cross-media learning approach is proposed and tested in this context. We assessed the validity of our model by conducting comparative studies and evaluations on a benchmark for visual sentiment analysis. Our empirical study shows that although the text associated to each image is often noisy and weakly correlated with the image content, it can be profitably exploited to train a deep Convolutional Neural Network that effectively predicts the sentiment polarity of previously unseen images. The dataset used in our experiments, named T4SA (Twitter for Sentiment Analysis), is available on this page.

Data collection process.

Paper (PDF will be linked here, X.XMB)

The paper will be presented at the 5th Workshop on Web-scale Vision and Social Media (VSM, 23rd October 2017), ICCV 2017
BibTeX will go here

Trained Models

Here you can download the trained models used in our experiments. Models have been trained with BVLC Caffe and DIGITS. The following archives can be directly imported into DIGITS.

vgg19_finetuned_all.tar.gz (518.2 MB)
vgg19_finetuned_fc6+.tar.gz (518.2 MB)
Hybrid-T4SA FT-A
hybrid_finetuned_all.tar.gz (210.9 MB)
Hybrid-T4SA FT-F
hybrid_finetuned_fc6+.tar.gz (210.9 MB)

Hybrid-T4SA models are Hybrid-CNNs (AlexNet trained on ILSVRC12 + Places205) finetuned on our T4SA dataset, while VGG-T4SA models are VGG-19 nets trained on ILSVRC12 and then finetuned on T4SA. In FT-A models, all layers have been finetuned, while in FT-F model convolutional layers are fixed and only fully-connected layers are finetuned.

You can check the performance of the available models in the Experimental Results section.


The data collection process took place from July to December 2016, lasting around 6 months in total. During this time span, we exploited Twitter's Sample API to access a random 1% sample of the stream of all globally produced tweets, discarding:

Each tweet (text and associated images) has been labeled according to the sentiment polarity of the text (positive, neutral and negative sentiment) predicted by our tandem LSTM-SVM architecture, obtaining a labeled set of tweets and images divided in 3 categories we called T4SA. We removed near-duplicate images and we selected a balanced subset of images, named B-T4SA, that we used to train our visual classifiers.

The details of the dataset are reported in the following table. You can request the access to the T4SA dataset using the form below.

(w/o near dups)
Positive 371,341 501,037 372,904 156,862
Neutral 629,566 757,895 444,287 156,862
Negative 179,050 214,462 156,862 156,862
Sum 904,395 1,473,394 974,053 470,586
Request Access to the Dataset

Experimental Results

In the first table, we report the performance of the models we trained, together with other previously published methods, on two testing datasets: Twitter Testing Dataset and B-T4SA test split.

In the second table, we report the accuracy of different models when performing cross-validation on the Twitter Testing Datatset.

Model Twitter Testing Dataset
(2-way: pos, neg)
B-T4SA test set
(3-way: pos, neu, neg)
5 agree ≥ 4 agree ≥ 3 agree
Random Classifier 50.0% 50.0% 50.0% 33.3%
CNN (You et al.) 72.2% 68.6% 66.7% -
PCNN (You et al.) 74.7% 71.4% 68.7% -
Hybrid-T4SA FT-F 76.6% 74.8% 72.3% 49.9%
Hybrid-T4SA FT-A 74.1% 70.9% 68.6% 49.1%
VGG-T4SA FT-F 76.8% 73.7% 71.5% 50.6%
VGG-T4SA FT-A 78.5% 75.5% 72.5% 51.3%
Prediction accuracy of various models. Our models have been trained on B-T4SA train set, then tested on the various splits of Twitter Testing Dataset and on our B-T4SA test set.
Twitter Testing Dataset
Method Training derails 5 agrees ≥ 4 agrees ≥ 3 agrees
Approaches without intermediate fine-tuning
GCH * - 68.4% 66.5% 66.0%
SentiBank - 70.9% 67.5% 66.2%
LCH * - 71.0% 67.1% 66.4%
GCH + BoW * - 71.0% 68.5% 66.5%
LCH + BoW * - 71.7% 69.7% 66.4%
Sentribute - 73.8% 70.9% 69.6%
CNN Custom architecture tr. on Flickr (VSO) 78.3% 75.5% 71.5%
AlexNet AlexNet tr. on ILSVRC2012 81.7% 78.2% 73.9%
PlaceCNN AlexNet tr. on Places205 83.0% - -
GoogleNet GoogleNet tr. on ILSVRC2012 86.1% 80.7% 78.7%
HybridNet AlexNet tr. on (ILSVRC2012 + Places205) 86.7% 81.4% 78.1%
VGG-19 VGG-19 tr. on ILSVRC2012 88.1% 83.5% 80.0%
Approaches using an intermediate fine-tuning
PCNN Custom architecture tr. on Flickr (VSO) + ft. on Flickr (VSO) 77.3% 75.9% 72.3%
DeepSentiBank ○● AlexNet tr. on ILSVRC2012 + ft. on Flickr (VSO) 80.4% - -
MVSO [EN] ○● DeepSentiBank ft. on MVSO-EN 83.9% - -
Hybrid-T4SA FT-A (Ours) AlexNet tr. on (ILSVRC2012 + Places205) + ft. on B-T4SA 86.4% 83.0% 80.0%
Hybrid-T4SA FT-F (Ours) AlexNet tr. on (ILSVRC2012 + Places205) + ft. on B-T4SA 87.3% 83.2% 81.0%
VGG-T4SA FT-F (Ours) VGG-19 tr. on ILSVRC2012 + ft. on B-T4SA 88.9% 85.7% 81.5%
VGG-T4SA FT-A (Ours) VGG-19 tr. on ILSVRC2012 + ft. on B-T4SA 89.6% 86.6% 82.0%

* Approch based on low-level features         Approch based on mid-level features         Approch based on deep learning

Five-fold cross-validation accuracy of different methods on the Twitter Testing Dataset. tr. stands for ‘trained’; ft. stands for ‘fine-tuned’. Note that in these experiments all the deep models are again fine-tuned on four folds of the Twitter Testing Dataset. During cross-validation we fine-tuned all the weights of our FT models.
The following figure shows the most confident classifications of our best model on our B-T4SA test set, grouped by all possible (groundtruth, predicted class) couples. Rows (from top to bottom) contains images labeled respectively negative, neutral and positive on the basis of textual sentiment analysis. Columns (from left to right) contain images visually classified respectively as negative, neutral and positive by our model.