UNOSAT-AI-Building-Footprint

This work been done by Shane at UNOSAT

The goal of UNOSAT-AI-Building-Footprint Project is to detect building on high resolution satellite imagery. It can get the geojson file of the corresponding satellite imagery end to end.

It consists of 4 parts:

1. Dataset Preparation: Convert Shpfile to coco format annotation

2. Detection Model: we use the DetectoRS as the detection model

3. The whole Satellite imagery inference strategy: Slice the whole satellite and convert the output of patches to the whole geojson file.

4. Shell scripts of the pipeline.

---

0. Install

---

We provide a Dockerfile to build an image for the pipline' environment.

# build an image with PyTorch 1.3, CUDA 10.1,CUDNN 7
docker build --no-cache -t unosataifootprint .

Run it with

nvidia-docker run -it --ipc=host -v {UNOSAT-Building-Footprint}:/workspace/Buildingfootprint unosataifootprint

---

1. Dateset preparation

---

dataset_preparation.sh

---

2. Train the detection model

---

cd /DetectoRS-master-UNOSAT
tools/dist_train.sh {CONFIGS} {GPUNUMS}

for example: tools/dist_train.sh configs/DetectoRS/DetectoRS_mstrain_400_1200_r50_40e.py 4

---

3. Inferrence the whole satellite imagery

Just need put the satellite imagery on the OrgImage directory

Run the shell below

---

infer_tot.sh

We can get the corresponding output geojson file on the Output directly

aws

UNOSAT-AI-Building-Footprint-ok

TO DO:

Multi-GPU test

dist_test.sh

infer_json_test.py

ann_file.py

open-mmlab/mmdetection#2167

Thanks to the three open source code

DetectoRS
simrdwn
shp2coco