Aerial Vehicle Detection using Synthetic Images
This repository contains our work on Aerial Vehicle Detection using Synthetic Images
. You can find more details on our work in the
following paper :
@article{uzkent2017tracking,
Title={Tracking in Aerial Hyperspectral Videos using Deep Kernelized Correlation Filters},
Author={Uzkent, Burak and Rangnekar, Aneesh and Hoffman, Matthew J},
Journal={arXiv preprint arXiv:1711.07235},
Year={2017}
}
Synthetic Training Dataset
The goal of our work is detect vehicles on the WAMI
platform by using convolutional model trained on a synthetic vehicle detection dataset. Our synthetic dataset consists of 55226
images with vehicle and background samples and each image is represented by 48x48
pixels. The synthetic images are generated by the Digital Imaging and Remote Sensing
Software (DIRSIG) owned by the Chester F. Carlson Center for Imaging Science, Rochester Institute of Technology
. The synthetic scenes comes from the MegaScene-I area available in DIRSIG.The vehicular traffic simulation, on ther other hand, is generated by the Simulation of Urban Mobility Platform
(SUMO) and coupled to DIRSIG. Some of the positive samples from our dataset is shown below.
Validation Dataset
To test the fidelity of the synthetic dataset, we validate the performance of the convolutional network, trained on the synthetic dataset, on the validation dataset containing the real vehicle and background images recorded with the WAMI
platform. Our validation dataset consists of 600
images from the CLIFF06
and CLIFF07
datasets. Some of the positive images from the validation dataset can be visualized in the figure below.
The goal of this study is to reduce the dependency on the WAMI dataset to train a convolutional network model. With our synthetic dataset, we believe that the overfitting
to the a couple of datasets captured from the WAMI dataset is avoided.
Download Links
Download our aerial vehicle detection dataset
wget https://drive.google.com/open?id=1cQIM2a7gNaxlE2oFdQ_O-GqgBo84fLia
If you use our aerial vehicle detection dataset in your research project, please cite our paper :
@article{uzkent2017tracking,
Title={Tracking in Aerial Hyperspectral Videos using Deep Kernelized Correlation Filters},
Author={Uzkent, Burak and Rangnekar, Aneesh and Hoffman, Matthew J},
Journal={arXiv preprint arXiv:1711.07235},
Year={2017}
}
Train ResNet50 Model on the DIRSIG Vehicle Classification Dataset
In this step, the DIRSIG
dataset is used to train the tensorflow implementation of the ResNet50
network. To train the model, first download the dataset as explained in the steps above. For the next step, please visit the following README file.