AF-SSD: An Accurate and Fast Single Shot Detector for High Spatial Remote Sensing Imagery

by Ruihong Yin, Wei Zhao, Xudong Fan, Yongfeng Yin

Introduction

Contents

• Installation
• Datasets
• Training
• Evaluation
• Demo
• Performance

Installation

• Install PyTorch 0.4.0 by the instrument on the website Pytorch and running the approriate command.
• Clone this repository. This repository is mainly based on ssd.pytorch, a huge thank to them.
  • Note: We currently only support Python 3+ and Pytorch 0.4.0.

Datasets

NWPU VHR-10 dataset is avalable here.

Training

• Pre-trained MobileNetv1 is downloaded from our BaiduYun Driver(code :h4y7). By default, the pre-trained module is in the AF-SSD/weights dir.
• To train AF-SSD using the train script to specify the parameters listed in train_AFSSD.py as a flag or manually change them:

python train_AFSSD.py

Evaluation

To evaluate a trained network:

python test_AFSSD.py --trained_model ./weights/AFSSD_VOC_60000.pth

Note: you can specify the parameters listed in the test_AFSSD.py file by flagging them or manually changing them.

Demo

To test an image with a trained network:

python demo/demo.py

Note: you can change the parameters listed in the file.

Performance

NWPU VHR-10

System	mAP	Average Running Time
COPD	54.6%	1.070s
YOLOv2	60.5%	0.026s
RICNN	72.6%	8.770s
R-P-Faster RCNN	76.5%	0.150s
NEOON	77.5%	0.059s
SSD*	80.5%	0.042s
Faster RCNN	80.9%	0.430s
CACMOD CNN	90.4%	2.700s
AF-SSD	88.7%	0.035s

Note:

• The result of SSD* is our reproduced result with the same parameters as AF-SSD.
• The testing environment is NVIDIA GTX-1080Ti.