Ana's Ping Pong

Data science retreat batch 28 (https://datascienceretreat.com/) portfolio project.

We use satellite images and computer vision to spot and tag public ping pong tables on a map.

Object detection model

We re-trained detectron2 model ‘faster_rcnn_R_101_FPN_3x’ (pretrained on COCO dataset) using approx. 1200 bounding-box labeled map tiles downloaded from bing maps. Model loading is done in webserver/anaspingpong/predictor.py.

Webserver

For the prediction, 25 tiles in the current center of the map are temporarily downloaded and bounding boxes of ping pong tables are predicted. The position of the bounding box center is then saved in a sqlite database and displayed on the map using a ping pong racket icon.

Life webserver

The life webserver is not yet public, but we'll make the URL available here soon

Install webserver locally

• get your own key for Google maps API
• create file webserver/instance/config.py with GOOGLE_MAPS_KEY = "your_key""
• create new environment: python3 -m venv venv
• activate: . venv/bin/activate
• install requirements: pip install -r requirements
• pip3 install torch==1.10.0+cpu torchvision==0.11.1+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
• python -m pip install 'git+https://github.com/facebookresearch/detectron2.git')
• cd webserver
• export FLASK_APP=anaspingpong
• export FLASK_ENV=development
• flask init-db
• flask run
• application is now running on http://127.0.0.1:5000