Skip to content
Self-supervised learning of visual features through embedding images into text topic spaces
Branch: master
Clone or download
yash0307 Merge pull request #6 from yash0307/master
Add Wikipedia dataset.
Latest commit 35c17d2 Aug 18, 2018
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
CNN/CaffeNet initial commit May 23, 2017
LDA Add gensim logger. Mar 27, 2018
data Update README.md Jul 3, 2017
experiments + sync Jul 4, 2017
wikipedia_data Update README.md Aug 18, 2018
README.md Update README.md Jul 10, 2017
texttopicnet.png initial commit May 23, 2017

README.md

TextTopicNet

Self-supervised learning of visual features through embedding images into text topic spaces, CVPR 2017.

L. Gomez*, Y. Patel*, M. Rusiñol, D. Karatzas, C.V. Jawahar.

In this paper we present a method that is able to take advantage of freely available multi-modal content to train computer vision algorithms without human supervision. We put forward the idea of performing self-supervised learning of visual features by mining a large scale corpus of multi-modal (text and image) documents. We show that discriminative visual features can be learnt efficiently by training a CNN to predict the semantic context in which a particular image is more probable to appear as an illustration. For this we leverage the hidden semantic structures discovered by a well-known topic modeling technique over the text corpus.

TextTopicNet diagram

Our CNN learns to predict the semantic context in which images appear as illustration. Given an illustrated article we project its textual information into the topic-probability space provided by the LDA topic modeling framework. Then we use this semantic level representation as the supervisory signal for CNN training.

If you make use of the code in this repository for scientific purposes, we appreciate it if you cite our paper:

@inproceedings{GomezPatelCVPR2017self,
  title     = {Self-supervised learning of visual features through embedding images into text topic spaces},
  author    = {Lluis Gomez, Yash Patel, Marçal Rusiñol, Dimosthenis Karatzas, and C.V. Jawahar},
  booktitle = {CVPR},
  year      = {2017}
}

Pretrained models

TextTopicNet_Wikipedia_ImageCLEF_40Topics.caffemodel

TextTopicNet_finetuned_VOC2007.caffemodel

Training from scratch

Set up data

Download and uncompress the Wikipedia Retrieval 2010 Collection and The PASCAL Visual Object Classes Challenge 2007 datasets into the data/ImageCLEF_Wikipedia/ and data/VOC2007/ folders respectively. Follow the specific instructions provided in:

data/ImageCLEF_Wikipedia/README.md

data/VOC2007/README.md

Learn LDA model

Run the following commands in order to process the data and learn the LDA model on the text corpus:

cd LDA
python generate_train_dict.py
python learn_LDA_model.py
python generate_train_labels.py

This will generate the LDA model files (ldamodel40.lda and ldamodel40.lda.state) that can be used to project new texts into the learned topic space; and a file training_labels40.json with the trainig labels (topic space probabilities) for all images in our training dataset. If you want to skip this step and directly train the CNN model you can download the files from the following links: ldamodel40.lda / ldamodel40.lda.state / dictionary.dict / training_labels40.json

Train CNNs

Run the following commands to train the CaffeNet model from scratch:

cd CNN/CaffeNet
export PYTHONPATH=./:$PYTHONPATH
caffe train --solver solver.prototxt

Once trained, the CNN model can be used for image classification in two different ways:

  • as a feature extractor, i.e. extract features from late layers (e.g. fc7, fc6, pool5, etc.) and training an SVM classifier.
  • finetuning the weights for a classification task, e.g. using the PASCAL VOC 2007 training data.

The same model can be used for the following multi-modal retrieval tasks:

  • Image query vs. Text database: First projecting all text documents into the topic space using the learned LDA model. And then projecting the image query into the topic space using the CNN model, and find the nearest neighbour texts on the database.
  • Text query vs. Image database: same as before but the other way around.

Experiments

The scripts in the experiments/ folder reproduce the results of the experiments in our paper.

  • Run the script experiments/multi_modal_retrieval.py for multi-modal retrieval experiment on Wikipedia dataset.
  • Run the script experiments/voc_2007_classification.py for image classification experiment on PASCAL VOC 2007.
You can’t perform that action at this time.