SCAF-Net

A Scene Context Attention-Based Fusion Network for Vehicle Detection

Installation

git clone https://github.com/minghuicode/SCAF-Net
cd SCAF-Net
conda env create -f environment.yml 

Dataset Downloads

downloads DLR-3K dataset at dlr.de

cd SCAF-Net/data
wget https://pba-freesoftware.eoc.dlr.de/MunichDatasetVehicleDetection-2015-old.zip
unzip MunichDatasetVehicleDetection-2015-old.zip
ln -sf MunichDatasetVehicleDetection-2015-old/Train dlr

Model Training

There are total 10 labeled aerial images. We use 5 of them for training, others for test.

cd SCAF-Net
conda activate torch
python train.py

Model Evaluation

To evaluate model performance on other 5 labeled aerial images, just run test file.

cd SCAF-Net
conda activate torch
python test.py --evaluation

Predict

To predict several unseen aerial images, run test files as follow. Visual output will be saved at output folder.

cd SCAF-Net
conda activate torch
mkdir input
cp data/dlr/*JPG input/
python test.py