Semantic Compositional Networks

The Theano code for the CVPR 2017 paper “Semantic Compositional Networks for Visual Captioning”

Model architecture and illustration of semantic composition.

Dependencies

This code is written in python. To use it you will need:

Python 2.7 (do not use Python 3.0)
Theano 0.7 (you can also use the most recent version)
A recent version of NumPy and SciPy

Getting started

We provide the code on how to train SCN for image captioning on the COCO dataset.

In order to start, please first download the ResNet features and tag features we used in the experiments. Put the coco folder inside the ./data folder.
We also provide our pre-trained model on COCO. Put the pretrained_model folder into the current directory.
In order to evaluate the model, please download the standard coco-caption evaluation code. Copy the folder pycocoevalcap into the current directory.
Now, everything is ready.

How to use the code

Run SCN_training.py to start training. On a modern GPU, the model will take one night to train.

THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python SCN_training.py

Based on our pre-trained model, run SCN_decode.py to generate captions on the COCO small 5k test set. The generated captions are also provided, named coco_scn_5k_test.txt.
Now, run SCN_evaluation.py to evaluate the model. The code will output

CIDEr: 1.043, Bleu-4: 0.341, Bleu-3: 0.446, Bleu-2: 0.582, Bleu-1: 0.743, ROUGE_L: 0.550, METEOR: 0.261.

In the ./data/coco folder, we also provide the features for the COCO official validation and test sets. Run SCN_for_test_server.py will help you generate captions for the official test set, and prepare the .json file for submission.

Demo: generating image tags and captions from scratch

We provide a demo code here that detects tags and generates captions from scratch. The code is for research purposes only. The code runs on Windows system and requires Python 3.6.

We provide two running modes. First, generate a caption from a given image url. Second, generate captions for a batch of images.

We can generate a caption from an image url by running demo_image_captioning.py. Below are two examples.

python demo_image_captioning.py --image_url http://c8.alamy.com/comp/CR318J/border-collie-playing-with-a-frisbee-in-its-mouth-on-a-meadow-with-CR318J.jpg

Outputs:

start image captioning @ 12:38:08.430527
first, downloading the image ...
second, extract image features ...
Now, start image captioning ...
Detected tags: grass (1.0), field (0.985), dog (0.977), outdoor (0.94), black (0.919), yellow (0.65), green (0.591), grassy (0.388), small (0.341), standing (0.301),
Generated captions: a black dog with a frisbee in its mouth
end @ 12:38:13.249350

Another example:

python demo_image_captioning.py --image_url https://coresites-cdn.factorymedia.com/mpora_new/wp-content/uploads/2017/10/Porthleven-Surf-Sep2017-%C2%A9EdBlomfield-100.jpg

Outputs:

start image captioning @ 12:38:16.257385
first, downloading the image ...
second, extract image features ...
Now, start image captioning ...
Detected tags: water (1.0), sport (0.997), surfing (0.997), wave (0.995), outdoor (0.993), riding (0.978), ocean (0.946), man (0.756), top (0.463), board (0.448),
Generated captions: a man riding a wave on top of a surfboard
end @ 12:38:21.217550

The two images corresponding to the image urls above are shown below.

We can also generate captions for a batch of images by running demo_image_captioning_batch.py.

python demo_image_captioning_batch.py

Outputs look like the following:

start image captioning @ 21:00:20.145487
First, extract image features ...
Now, start image captioning ...
Image name: COCO_test2014_000000000001.jpg
Detected tags: outdoor (0.999), grass (0.996), truck (0.973), fence (0.97), road (0.872), street (0.515), car (0.494), green (0.466), next (0.438), parked (0.391),
Generated captions: a truck is parked on the side of the road

Image name: COCO_test2014_000000000014.jpg
Detected tags: road (1.0), outdoor (0.994), building (0.973), street (0.927), person (0.811), motorcycle (0.705), man (0.609), sidewalk (0.585), walking (0.443), crossing (0.389),
Generated captions: a man riding a motorcycle down a street

...

end @ 21:00:45.461410

Generating captions for 10 images roughly takes 25 seconds.

Video Captioning

In order to keep things simple, we provide another separate repo that reproduces our results on video captioning, using the Youtube2Text dataset.

Citing SCN

Please cite our CVPR paper in your publications if it helps your research:

@inproceedings{SCN_CVPR2017,
  Author = {Gan, Zhe and Gan, Chuang and He, Xiaodong and Pu, Yunchen and Tran, Kenneth and Gao, Jianfeng and Carin, Lawrence and Deng, Li},
  Title = {Semantic Compositional Networks for Visual Captioning},
  booktitle={CVPR},
  Year  = {2017}
}

zhegan27 / Semantic_Compositional_Nets