Want to take your software engineering career to the next level? Join the mailing list for career tips & advice Click here


Dense image captioning in Torch

Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402Star full 4f7b624809470f25b6493d5a7b30d9b9cb905931146e785d67c86ef0c205a402 (1 ratings)
Rated 5.0 out of 5
Subscribe to updates I use densecap

Statistics on densecap

Number of watchers on Github 1012
Number of open issues 37
Average time to close an issue 12 days
Main language Jupyter Notebook
Average time to merge a PR about 2 hours
Open pull requests 1+
Closed pull requests 2+
Last commit about 3 years ago
Repo Created over 4 years ago
Repo Last Updated over 2 years ago
Size 8.85 MB
Organization / Authorjcjohnson
Page Updated
Do you use densecap? Leave a review!
View open issues (37)
View densecap activity
View on github
Fresh, new opensource launches 🚀🚀🚀
Software engineers: It's time to get promoted. Starting NOW! Subscribe to my mailing list and I will equip you with tools, tips and actionable advice to grow in your career.
Evaluating densecap for your project? Score Explanation
Commits Score (?)
Issues & PR Score (?)


This is the code for the paper

DenseCap: Fully Convolutional Localization Networks for Dense Captioning,
Justin Johnson*, Andrej Karpathy*, Li Fei-Fei,
(* equal contribution)
Presented at CVPR 2016 (oral)

The paper addresses the problem of dense captioning, where a computer detects objects in images and describes them in natural language. Here are a few example outputs:

The model is a deep convolutional neural network trained in an end-to-end fashion on the Visual Genome dataset.

We provide:

If you find this code useful in your research, please cite:

  title={DenseCap: Fully Convolutional Localization Networks for Dense Captioning},
  author={Johnson, Justin and Karpathy, Andrej and Fei-Fei, Li},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and 
             Pattern Recognition},


DenseCap is implemented in Torch, and depends on the following packages: torch/torch7, torch/nn, torch/nngraph, torch/image, lua-cjson, qassemoquab/stnbhwd, jcjohnson/torch-rnn

After installing torch, you can install / update these dependencies by running the following:

luarocks install torch
luarocks install nn
luarocks install image
luarocks install lua-cjson
luarocks install https://raw.githubusercontent.com/qassemoquab/stnbhwd/master/stnbhwd-scm-1.rockspec
luarocks install https://raw.githubusercontent.com/jcjohnson/torch-rnn/master/torch-rnn-scm-1.rockspec

(Optional) GPU acceleration

If have an NVIDIA GPU and want to accelerate the model with CUDA, you'll also need to install torch/cutorch and torch/cunn; you can install / update these by running:

luarocks install cutorch
luarocks install cunn
luarocks install cudnn

(Optional) cuDNN

If you want to use NVIDIA's cuDNN library, you'll need to register for the CUDA Developer Program (it's free) and download the library from NVIDIA's website; you'll also need to install the cuDNN bindings for Torch by running

luarocks install cudnn

Pretrained model

You can download a pretrained DenseCap model by running the following script:

 sh scripts/download_pretrained_model.sh

This will download a zipped version of the model (about 1.1 GB) to data/models/densecap/densecap-pretrained-vgg16.t7.zip, unpack it to data/models/densecap/densecap-pretrained-vgg16.t7 (about 1.2 GB) and then delete the zipped version.

This is not the exact model that was used in the paper, but is has comparable performance; using 1000 region proposals per image, it achieves a mAP of 5.70 on the test set which is slightly better than the mAP of 5.39 that we report in the paper.

Running on new images

To run the model on new images, use the script run_model.lua. To run the pretrained model on the provided elephant.jpg image, use the following command:

th run_model.lua -input_image imgs/elephant.jpg

By default this will run in GPU mode; to run in CPU only mode, simply add the flag -gpu -1.

This command will write results into the folder vis/data. We have provided a web-based visualizer to view these results; to use it, change to the vis directory and start a local HTTP server:

cd vis
python -m SimpleHTTPServer 8181

Then point your web browser to http://localhost:8181/view_results.html.

If you have an entire directory of images on which you want to run the model, use the -input_dir flag instead:

th run_model.lua -input_dir /path/to/my/image/folder

This run the model on all files in the folder /path/to/my/image/folder/ whose filename does not start with ..

The web-based visualizer is the prefered way to view results, but if you don't want to use it then you can instead render an image with the detection boxes and captions baked in; add the flag -output_dir to specify a directory where output images should be written:

th run_model.lua -input_dir /path/to/my/image/folder -output_dir /path/to/output/folder/

The run_model.lua script has several other flags; you can find details here.


To train a new DenseCap model, you will following the following steps:

  1. Download the raw images and region descriptions from the Visual Genome website
  2. Use the script preprocess.py to generate a single HDF5 file containing the entire dataset (details here)
  3. Use the script train.lua to train the model (details here)
  4. Use the script evaluate_model.lua to evaluate a trained model on the validation or test data (details here)

For more instructions on training see INSTALL.md in doc folder.


In the paper we propose a metric for automatically evaluating dense captioning results. Our metric depends on METEOR, and our evaluation code requires both Java and Python 2.7. The following script will download and unpack the METEOR jarfile:

sh scripts/setup_eval.sh

The evaluation code is not required to simply run a trained model on images; you can find more details about the evaluation code here.

Webcam demos

If you have a powerful GPU, then the DenseCap model is fast enough to run in real-time. We provide two demos to allow you to run DenseCap on frames from a webcam.

Single-machine demo

If you have a single machine with both a webcam and a powerful GPU, then you can use this demo to run DenseCap in real time at up to 10 frames per second. This demo depends on a few extra Lua packages:

You can install / update these dependencies by running the following:

luarocks install camera
luarocks install qtlua

You can start the demo by running the following:

qlua webcam/single_machine_demo.lua

Client / server demo

If you have a machine with a powerful GPU and another machine with a webcam, then this demo allows you use the GPU machine as a server and the webcam machine as a client; frames will be streamed from the client to to the server, the model will run on the server, and predictions will be shipped back to the client for viewing. This allows you to run DenseCap on a laptop, but with network and filesystem overhead you will typically only achieve 1 to 2 frames per second.

The server is written in Flask; on the server machine run the following to install dependencies:

cd webcam
virtualenv .env
source .env/bin/activate
pip install -r requirements.txt
cd ..

For technical reasons, the server needs to serve content over SSL; it expects to find SSL key files and certificate files in webcam/ssl/server.key and webcam/ssl/server.crt respectively. You can generate a self-signed SSL certificate by running the following:

mkdir webcam/ssl

# Step 1: Generate a private key
openssl genrsa -des3 -out webcam/ssl/server.key 1024
# Enter a password

# Step 2: Generate a certificate signing request
openssl req -new -key webcam/ssl/server.key -out webcam/ssl/server.csr
# Enter the password from above and leave all other fields blank

# Step 3: Strip the password from the keyfile
cp webcam/ssl/server.key webcam/ssl/server.key.org
openssl rsa -in webcam/ssl/server.key.org -out webcam/ssl/server.key

# Step 4: Generate self-signed certificate
openssl x509 -req -days 365 -in webcam/ssl/server.csr -signkey webcam/ssl/server.key -out webcam/ssl/server.crt
# Enter the password from above

You can now run the following two commands to start the server; both will run forever:

th webcam/daemon.lua
python webcam/server.py

On the client, point a web browser at the following page:


but you should replace SERVER_URL with the actual URL of the server.

Note: If the server is using a self-signed SSL certificate, you may need to manually tell your browser that the certificate is safe by pointing your client's web browser directly at the server URL; you will get a message that the site is unsafe; for example on Chrome you will see the following:

Afterward you should see a message telling you that the DenseCap server is running, and the web client should work after refreshing.

densecap open issues Ask a question     (View All Issues)
  • over 3 years Problem when running webcam client/server demo
  • over 3 years Small stumbling stones, and ideas for readme
  • over 3 years Warning: cudnn.convert does not work with nngraph yet. Ignoring nn.gModule
  • over 3 years which IDE do you use to debug lua?
  • over 3 years cuda runtime error : too many resources requested for launch
  • almost 4 years how to train the model if we don't have ground truth for the bounding box
  • almost 4 years Big charge of cpu
  • almost 4 years Help the preprocess.py error!
  • almost 4 years Confused by the METEOR metric in paper
  • almost 4 years some errors when running train.lua
  • almost 4 years ./densecap/modules/BoxSampler.lua:105: attempt to call method 'scatter' (a nil value)
  • almost 4 years module 'sys.fpath' not found:No LuaRocks module found for sys.fpath
  • almost 4 years How can I use natural language queries to retrieve the source image?
  • almost 4 years Some minor bugs
  • about 4 years Does training require GPU?
  • about 4 years Backward pass
  • about 4 years Failing Tests
densecap open pull requests (View All Pulls)
  • Update INSTALL.md
densecap list of languages used
Other projects in Jupyter Notebook