This is an official Tensorflow implementation of single object tracking in videos by using hierarchical attentive recurrent neural networks, as presented in the following paper:

A. R. Kosiorek, A. Bewley, I. Posner, "Hierarchical Attentive Recurrent Tracking", NIPS 2017.

• Author: Adam Kosiorek, Oxford Robotics Institue, University of Oxford
• Email: adamk(at)robots.ox.ac.uk
• Paper: https://arxiv.org/abs/1706.09262
• Webpage: http://ori.ox.ac.uk/

Install Tensorflow v1.1 and the following dependencies (using pip install -r requirements.txt (preferred) or pip install [package]):

• matplotlib==1.5.3
• numpy==1.12.1
• pandas==0.18.1
• scipy==0.18.1

The notebook scripts/demo.ipynb contains a demo, which shows how to evaluate tracker on an arbitrary image sequence. By default, it runs on images located in imgs folder and uses a pretrained model. Before running the demo please download AlexNet weights first (described in the Training section).

1. Download KITTI dataset from here. We need left color images and tracking labels.
2. Unpack data into a data folder; images should be in an image folder and labels should be in a label folder.
3. Resize all the images to (heigh=187, width=621) e.g. by using the scripts/resize_imgs.sh script.

1. Download the AlexNet weights:

   • Execute scripts/download_alexnet.sh or
   • Download the weights from here and put the file in the checkpoints folder.

2. Run

    python scripts/train_hart_kitti.py --img_dir=path/to/image/folder --label_dir=/path/to/label/folder
   

The training script will save model checkpoints in the checkpoints folder and report train and test scores every couple of epochs. You can run tensorboard in the checkpoints folder to visualise training progress. Training should converge in about 400k iterations, which should take about 3 days. It might take a couple of hours between logging messages, so don't worry.

The scripts/eval_kitti.ipynb notebook contains the code necessary to prepare (IoU, timesteps) curves for train and validation set of KITTI. Before running the evaluation:

• Download AlexNet weights (described in the Training section).
• Update image and label folder paths in the notebook.

If you find this repo useful in your research, please consider citing:

@inproceedings{Kosiorek2017hierarchical,
   title = {Hierarchical Attentive Recurrent Tracking},
   author = {Kosiorek, Adam R and Bewley, Alex and Posner, Ingmar},
   booktitle = {Neural Information Processing Systems},
   url = {http://www.robots.ox.ac.uk/~mobile/Papers/2017NIPS_AdamKosiorek.pdf},
   pdf = {http://www.robots.ox.ac.uk/~mobile/Papers/2017NIPS_AdamKosiorek.pdf},
   year = {2017},
   month = {December}
}

This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/.

Version 1.0

• Original version from the paper. It contains the KITTI tracking experiment.