Real-time Vehicle Detection from UAV Imagery

By Xuemei Xie, Wenzhe Yang, Guimei Cao, etc.

Introduction

We proposed a novel UAV vehicle detector. The network is based on Refinedet, we improve it by generating more suitable bounding boxes and we propose a dynamic training strategy, which can help our model achieve the state-of-art accuracy on UAV dataset detection. More details please refer to the paper.

System	UAV dataset test mAP	FPS
YOLO	53.7	64
SSD300	83.3	59
Refinedet	87.1	40
Our preposed	90.8	59

Contents

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

Installation

1. Get the code. We will call the cloned directory as $caffe.

git clone https://github.com/wnzhyee/bigmm2018-code.git

2. Build the code. Please follow Caffe instruction to install all necessary packages and build it.

cd $caffe
# Modify Makefile.config according to your Caffe installation.
# Make sure to include $caffe/python to your PYTHONPATH.
cp Makefile.config.example Makefile.config
make all -j && make py

Preparation

1. Download UAV dataset. The dataset is prepared in VOC2007 style, and we provide the image files in ‘JPEGImages/’, the annotations files are in 'Annotations/'.

2. Create .lmdb files uses the create_list and create_data file.

Training

Train your model on PASCAL VOC. First change the path in DJI_train.sh shell and solver.prototxt(they are in $caffe/models/VGGNet/VOC0712/method_folder/), then run the train shell.

cd $caffe/models/VGGNet/VOC0712/**method_folder**/
sh DJI_train.sh

The method folder:

1.Only CE: only train the model with cross-entropy loss, and the total iterations are 120k;

2.Only FL: only train the model with focal loss, and the total iterations are 120k;

3.DTS: train 100k iters with CE loss and another 20k iters with FL loss;

Evaluation

The evaluation files are in $caffe/examples, please refer to the evaluation shells for image det and vedio det.