From Captions to Visual Concepts and Back

Code for detecting visual concepts in images.

Installation Instructions

0. Create directory, checkout code, caffe, coco-hooks

git clone git@github.com:s-gupta/visual-concepts.git code
git clone git@github.com:pdollar/coco.git code/coco
cd code/coco
git checkout 3736a0068b6e4634563b0c5847d7783dae9bd461
cd ../../

0. Make caffe and pycaffe

git clone git@github.com:s-gupta/caffe.git code/caffe 
cd code/caffe 
git checkout mil
make -j 16
make pycaffe
cd ../../

0. Get the COCO images, caffe imagenet models, pretrained models on COCO.

# Get the COCO splits, ground truth
wget ftp://ftp.cs.berkeley.edu/pub/projects/vision/im2cap-cvpr15b/data.tgz && tar -xf data.tgz

# Get the COCO images
cd code

# Download and unzip the images 
bash scripts/script_download_coco.sh

# The code assumes images to be stored heirarchically. 
# This python scripts does the required copying.
PYTHONPATH='.' python scripts/script_download_coco.py

# Get the caffe imagenet models 
wget ftp://ftp.cs.berkeley.edu/pub/projects/vision/im2cap-cvpr15b/caffe-data.tgz && tar -xf caffe-data.tgz

# Get the pretrained models. The precomputed results in the previous tarball
# were not complete. v2 tar ball fixes this.
wget ftp://ftp.cs.berkeley.edu/pub/projects/vision/im2cap-cvpr15b/trained-coco.v2.tgz && tar -xf trained-coco.v2.tgz 

Demo

Look at demo.ipynb for a demo in IPython notebook.

Training, Testing the model

cd code and execute relevant commands from the file scripts/scripts_all.py

Citing

If you find this codebase useful in your research, please consider citing the following paper:

@InProceedings{Fang_2015_CVPR,
  author = {Fang, Hao and Gupta, Saurabh and Iandola, Forrest and Srivastava, Rupesh K. and Deng, Li and Dollar, Piotr and Gao, Jianfeng and He, Xiaodong and Mitchell, Margaret and Platt, John C. and Lawrence Zitnick, C. and Zweig, Geoffrey},
  title = {From Captions to Visual Concepts and Back},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2015}
}