PocketFinder

Automated Finder for Pocket billiards venues.

Challenges:

1. Ambiguity in text
2. Limited pixels on object

Finding bars, resturants, and hotels with pool tables is particuarly hard problem because these places often don't list billiards or pool tables as an amentiy on their websites. Open Street Map (OSM) does have a "sport=billiards" tag that can be added to Points of Interest, however without a OSM-enthustist pool player in the area, venues are rarely mapped.

Text: Sometimes Google reviewers will include a note in their feedback, however there is high ambiguity between "pool tables", "pool", and "table". Eg "I really enjoyed playing in the pool, and got a table right away at the resturant." Rarely will a reviewer use "billiards" for a dive bar, unless it's a pool hall or higher-scale establishment.

Images: Scanning the photos, pool tables often show as small slivers in the back of full-room shots, making identifying the billiards table difficult.

-> Text Processing: TODO

-> Image Processing: This repo approaches the problem using an image segmentation model of a pre-trained YOLOv8, fine-tuned on pool tables from OpenImages.

The developers of PocketFinder would love for users to add the pool tables found to OSM with the "sport=billiards" tag. Query through Overpass Turbo API, or phone apps.
https://www.openstreetmap.org/ https://overpass-turbo.eu/#

Image Gathering and Inference Integration

python3 PocketFinder.py 

1. lat/lon/radius -> GooglePlacesVenueFinder -> place IDs
2. place IDs -> GooglePlacesPhotoReviews -> images + reviews
3. images -> PoolTableInference -> confidence: pool table/no pool table

TODO: reviews -> ReviewInference -> confidence: pool table/no pool table

Training + Inference

Setting up a Running Environment

Set up a working environmnet based off the requirements.txt. Using conda or miniconda.

Recommended order:

conda create --name pool python=3.10
conda activate pool
conda install pytorch pandas
pip install -r requirements.txt

Downloading the pool table images from openImages (and no pool tables)

python3 openimages-pool-table-image-downloader.py 
python3 negative-dataset-downloader.py 

(Optional) Check the images downloaded properly

python3 openimages-pool-checker.py 

Convert Openimages to YOLO format

python3 openimages-to-yolo-class.py

Config

Modify for your path.

python3 config.py

Fine-tuning YOLO for pool table recognition

python3 fine-tune-yolo-mps.py

Inference on New Single Images

python3 PoolTableInference.py

Note: Sections of this code was generated with Claude AI.