suniique / leveraging-crowdsourced-gps-data-for-road-extraction-from-aerial-imagery

The source code of CVPR 2019 paper "Leveraging Crowdsourced GPS Data for Road Extraction from Aerial Imagery"

Leveraging Crowdsourced GPS Data for Road Extraction from Aerial Imagery

The source code of CVPR 2019 paper "Leveraging Crowdsourced GPS Data for Road Extraction from Aerial Imagery"

Usage

python train.py \
	--model "dlink34" \
	--sat_dir "dataset/train_val/image" \
	--mask_dir "dataset/train_val/mask"\
	--gps_dir "dataset/GPS/patch" \
	--gps_type "data"
	

Dataset

Our dataset is avaliable upon request. Please contact via: suntao [AT] tongji.edu.cn

Dataset Description

• train_val/
  • image/: contains 278 satellite images (x_y_sat.png )
  • mask/: contains 278 mask images (x_y_mask.png )
• test/
  • image/: contains 70 satellite images (x_y_sat.png )
  • mask/: contains 70 mask images (x_y_mask.png )
• GPS/
  • beijing_gps_dir_speed_interval_sorted.pkl: The pickle file storing all raw GPS records
  • patch/: contains 348 GPS patch files (x_y_gps.pkl). Each stores the GPS records located in the area of input image x_y_sat.png
• coordinate/: contains x_y_coor.txt (WGS format) and x_y_coor2.txt (GCJ format) files

Each input image image/x_y_sat.png is a RGB satellite image of 1024 $\times$ 1024 pixel size. Its corresponding GPS data is stored in file /GPS/patch/x_y_gps.pkl, and corresponding mask image is mask/x_y_mask.png.

Unfortunately, we haven't got the permission to publish the satellite images due to the license of the data provider. However, we provide all the GPS coordinates of each satellite image (avaliable in WSG and GCJ format) in /coordinate/. You might apply for the access and download these images from Amap (高德地图) or DigitalGlobe referencing the coordinates.

GPS Data

The GPS dataset contains ~50m rows of GPS record collected from ~28k vehicles in Beijing.

To save the loading time, we publish the dataset in Python's Pickle format, which can be directly loaded like:

import pandas
import pickle
gps_data = pickle.load(open('beijing_gps_dir_speed_interval_sorted.pkl', 'rb'))

Here are first lines of this file:

	ID	time	lat	lon	dir	speed	timeinterval
0	0	1228061046	39.71743	116.61815	0	0	NaN
1	0	1228088457	39.71742	116.61798	0	0	177.5
2	0	1228088520	39.71670	116.61420	159	0	150.5
3	0	1228088758	39.71742	116.61798	0	0	272.5
4	0	1228090926	39.71670	116.61428	0	0	354.5
5	0	1228091249	39.73902	116.60902	12	308	318.0
6	0	1228091562	39.73770	116.56821	267	1080	264.0

Definition of columns:

• ID: Vehical ID (integer)
• time: Timestamp (UNIX format, in second)
• lat: Latitude (in degree)
• lon: Lontitude (in degree)
• dir: Heading (in degree, 0 means the vehical is heading north or isn't moving)
• speed: Speed (in meter per minute)
• timeinterval: The time interval between two records (in second)

The lat/lon are in the WGS System. The data table is sorted by ID and then by time.

License

This dataset is published under CC BY-NC-SA (Attribution-NonCommercial-ShareAlike) License . Please note that it can be ONLY used for academic or scientific purpose.

Citation

We kindly remind you that if found the code or dataset is useful for your research, please cite our paper.

Sun, Tao, Zonglin Di, Pengyu Che, Chun Liu, and Yin Wang. 
"Leveraging Crowdsourced GPS Data for Road Extraction From Aerial Imagery" 
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7509-7518. 2019.

leveraging-crowdsourced-gps-data-for-road-extraction-from-aerial-imagery's Issues

How to train the model without GPS data?

你好~

阅读完您的文章，我受益匪浅。
我有一个问题想咨询一下，如果我没有GPS数据，但仍想训练模型，请问在你的方法中我是否可以直接利用图片进行训练而不需要GPS数据？如果可以，我可以改动哪里呢？
多谢~

数据集获取

您好，我是一名硕士研究生，目前的方向是地物信息提取。您的论文_Leveraging-Crowdsourced-GPS-Data-for-Road-Extraction-from-Aerial-Imagery_给了我很大的启发。我希望能够使用该论文的数据集来深入并扩展我当前的研究。

如果可以，我希望能够获取包含卫星图像坐标、掩模和 GPS 数据的完整数据集，我的邮箱是[email protected]；[email protected]。
我愿意遵守数据使用的条件和相关隐私政策。感谢您考虑我的请求，期待您的回复。

Road dataset

Dear scholars, I am a master's student, and my current research direction is remote sensing image road extraction. Your paper "Leveraging-Crowdsourced-GPS-Data-for-Road-Extraction-from-Aerial-Imagery" has been of great help to me, and I would like to refer to the image and trajectory dataset of this paper to expand my research, if possible, my email is [[email protected]]
I am willing to comply with the conditions of data use and the relevant privacy policy. Thank you for considering my request and look forward to your response.

数据集请求

您好作者，我是一名武大做路网提取的研究生，读了您的文章，深受启发，可不可以共享下您本篇文章遥感影像与轨迹原始数据，我的邮箱是[email protected] 本人将不胜感激。

数据集请求

尊敬的学者，您好，我是一名硕士研究生，目前研究方向为遥感影像道路提取。您的论文《Leveraging-Crowdsourced-GPS-Data-for-Road-Extraction-from-Aerial-Imagery》给予我很大帮助，我希望参考该论文的影像与轨迹数据集来拓展研究，若准许，我的邮箱是[email protected]
我愿意遵守数据使用的条件和相关隐私政策。感谢您考虑我的请求，期待您的回复。

How to use multiple GPS features?

Thank you for your source code. I wonder what's the command to train with bearing layer and time interval layer as extra layers to GPS data, as well as GPS augmentation?

Is "1Dencoder" useful?

我在论文中好像没有看到关于1D encoder的实验，请问您实验1D encoder的效果如何呢？我能想到的一个问题是这个encoder没有Imagenet pretrined weights，可能需要自己先训。

数据集获取

尊敬的学者，您好，我是一名硕士研究生，目前研究方向为遥感影像道路提取。您的论文《Leveraging-Crowdsourced-GPS-Data-for-Road-Extraction-from-Aerial-Imagery》给予我很大帮助，我希望参考该论文的影像与轨迹数据集来拓展研究，若准许，我的邮箱是[[email protected]]
我愿意遵守数据使用的条件和相关隐私政策。感谢您考虑我的请求，期待您的回复。