VAID_dataset

Introduction

VAID(Vehicle Aerial Imaging from Drone) dataset contains 6000 aerial images under different illumination conditions, viewing angles from different places in Taiwan.The images are taken with the resolution of 1137 * 640 pixels in JPG format. Our VAID fataset contains seven classes of Vehicles, namely 'sedan' , 'minibus' , 'truck' , 'pickup truck' , 'bus' , 'cement truck' , ' trailer'.
From left to right in the figure below are labels 'sedan' , 'minibus' , 'truck' , 'pickup truck' , 'bus' , 'cement truck' , ' trailer' You can download VAID dataset in this website.https://vision.ee.ccu.edu.tw/aerialimage/

Test model

We test VAID dataset on 5 common model including Faster R-CNN, Yolov4, MobileNetv3 , RefineDet and U-Net.

How to use

Faster R-CNN

environment : tensorflow1.4 cuda9.2

1. Download the file https://drive.google.com/drive/folders/1S35gyY4c2B6HGzgB8WRl_scDYWdukhTb?usp=sharing
2. Put for_download/faster rcnn/data and for_download/faster rcnn/output into VAID_dataset/tf-faster-rcnn
3. Download dataset and put the fold Annotations amd JPEGImages into VAID_dataset/tf-faster-rcnn/data/VOCdevkit2007/VOC2007
4. make at VAID_dataset/tf-faster-rcnn/lib and VAID_dataset/tf-faster-rcnn/data/coco/PythonAPI

command :
demo
GPU_ID=0
CUDA_VISIBLE_DEVICES=${GPU_ID} ./tools/demo.py --net res101 --dataset pascal_voc

train
./experiments/scripts/train_faster_rcnn.sh 0 pascal_voc res101

test
./experiments/scripts/test_faster_rcnn.sh 0 pascal_voc res101

Yolov4

environment : cuda10.0 , cudnn>7.0 , opencv>2.4

1. Download the file https://drive.google.com/drive/folders/1S35gyY4c2B6HGzgB8WRl_scDYWdukhTb?usp=sharing
2. Put the folders for_download/yolov4/3rdparty , for_download/yolov4/backup , for_download/yolov4/cfd , for_download/yolov4/data , for_download/yolov4/include and for_download/yolov4/src into VAID_dataset/Yolov4/darknet
3. make in VAID_dataset/Yolov4/darknet
4. ./darknet if you succeed to make ,you will see usage: ./darknet
5. put for_download/yolov4/yolov4.conv.137 into VAID_dataset/Yolov4/darknet 6.Download dataset and put the fold Annotations amd JPEGImages into VAID_dataset-master/Yolov4/darknet/VOCdevkit VOC2007 and VAID_dataset-master/Yolov4/darknet/VOCdevkit VOC2007_test

command:
train
./darknet detector train cfg/voc.data cfg/yolo-obj.cfg yolov4.conv.137 calculate
map
./darknet detector map cfg/voc.data cfg/yolo-obj.cfg backup/yolo-obj_best.weights
test
./darknet detector test cfg/voc.data cfg/yolo-obj.cfg backup/yolo-obj_best.weights
If you want to know more command , you can see the official yolo github.

Mobilenetv3

environment : pytorch 1.4.0 cuda9.2 Download dataset and put the fold Annotations amd JPEGImages into VAID_dataset-master/MobileNetV3-SSD-Compact-Version/VAID run create_data_lists.py

command
python train.py
python eval.py

RefineDet

environment : cuda9.2 , pytorch 1.4.0

1. modify home path in VAID_dataset-master/RefineDet.PyTorch-master/data/config.py
2. modify VOC_ROOT path in VAID_dataset-master/RefineDet.PyTorch-master/data/voc0712.py
3. Download dataset and put the fold Annotations amd JPEGImages into VAID_dataset-master/RefineDet.PyTorch-master/data/VAID
4. Download the file https://drive.google.com/drive/folders/1S35gyY4c2B6HGzgB8WRl_scDYWdukhTb?usp=sharing
5. put for_download/Refinedet/weights into VAID_dataset-master/RefineDet.PyTorch-master

command:
train
./train_refinedet320.sh
calculate map
./eval_refinedet.sh

U-Net

environment : cuda 10.0 , Keras 2.2.4

1.Download the file https://drive.google.com/drive/folders/1S35gyY4c2B6HGzgB8WRl_scDYWdukhTb?usp=sharing 2.Put for_download/U-Net/data and for_download/U-Net/model into VAID_dataset/U-Net/ ; put for_download/U-Net/results into VAID_dataset/U-Net/MAP/ ; put for_download/U-Net/detection-results into VAID_dataset/U-Net/MAP/car/

3.Train(unet+_residual_vgg16.py) (1)modify 'path' to VAID_dataset/U-Net/data/train_val(seperate)/ (2)modify the model name you want to save in 'model_checkpoint'

4.Test(test.py) (1)modify 'test_path' to VAID_dataset/U-Net/data/test/test_jpg/ (2)modify 'label_save_dir' to VAID_dataset/U-Net/data/predict/predict_label/ (3)modify 'visualize_save_dir' to VAID_dataset/U-Net/data/predict/predict_visualize/ (4)modify 'modelFile' to model file(model_name.hdf5) in 'model' folder

5.IOU(evaluate_iou.py) (1)modify 'test_label_path' to VAID_dataset/U-Net/data/test/test_label/ (2)modify 'predict_path' to VAID_dataset/U-Net/data/predict/predict_label/

6.mAP(main_map.py)

Source

The code are from the following website.
Faster R-CNN : https://github.com/endernewton/tf-faster-rcnn
Yolov4 : https://github.com/AlexeyAB/darknet
Mobilenet : https://github.com/shaoshengsong/MobileNetV3-SSD-Compact-Version
RefineDet : https://github.com/luuuyi/RefineDet.PyTorch