Joint Event Detection and Description in Continuous Video Streams

Code released by Huijuan Xu (Boston University).

Introduction

We present the Joint Event Detection and Description Network (JEDDi-Net) that solves the dense captioning task in an end-to-end fashion. Our model continuously encodes the input video stream with three-dimensional convolutional layers, proposes variable-length temporal events based on pooled features, and transcribes the event proposals into captions with the consideration of visual and language context.

License

JEDDi-Net is released under the MIT License (refer to the LICENSE file for details).

Citing JEDDi-Net

If you find JEDDi-Net useful in your research, please consider citing:

@article{xu2019joint,
title={Joint Event Detection and Description in Continuous Video Streams},
author={Xu, Huijuan and Li, Boyang and Ramanishka, Vasili and Sigal, Leonid and Saenko, Kate},
journal={2019 IEEE Winter Conference on Applications of Computer Vision (WACV)},
year={2019}
}

Contents

1. Installation
2. Preparation
3. Training
4. Testing

Installation:

1. Clone the JEDDi-Net repository.

   git clone --recursive git@github.com:VisionLearningGroup/JEDDi-Net.git

2. Build Caffe3d with pycaffe (see: Caffe installation instructions).

   Note: Caffe must be built with Python support!

cd ./caffe3d

# If have all of the requirements installed and your Makefile.config in place, then simply do:
make -j8 && make pycaffe

3. Build JEDDi-Net lib folder.

   cd ./lib    
   make

Preparation:

1. Download the ground truth annatations and videos in ActivityNet Captions dataset.

2. Extract frames from downloaded videos in 25 fps.

3. Generate the pickle data for training and testing JEDDi-Net model.

   cd ./preprocess
   # generate training data
   python generate_train_roidb_sorted.py
   # generate validation data
   python generate_val_roidb.py  

Training:

1. Download the separately-trained segment proposal network(SPN) and captioning models ./pretrain/ .

2. In JEDDi-Net root folder, run:

   bash ./experiments/denseCap_jeddiNet_end2end/script_train.sh

Testing:

1. Download one sample JEDDi-Net model to ./snapshot/ .

   One JEDDi-Net model on ActivityNet Captions dataset is provided in: caffemodel .

   The provided JEDDi-Net model has the METEOR score ~8.58% on the validation set.

2. In JEDDi-Net root folder, generate the prediction log file on the validation set.

   bash ./experiments/denseCap_jeddiNet_end2end/test/script_test.sh 

3. Generate the results.json file from the prediction log file.

   cd ./experiments/denseCap_jeddiNet_end2end/test/
   bash bash.sh

4. Follow the evaluation code to get the evaluation results.