Foosball Dataset

This repository contains labeled images of various foosball tables.

Tables (so far):

• leonhart
• vector

Data Organization

This dataset is organized by manufacturer. Every manufacturer has its own set of figures and color-scheme.

The source of the data must be provided.

.
├── leonhart # the table manufacturer
│   └── black-green # flavor of table (e.g. color-scheme or specific model)
│       └── 01 # dataset number for this manufacturer/table combination
│           ├── pictures # contains the sourced pictures
│           └── out # contains the label information
│               └── YOLO_darknet
├── vector # table
├── vector #
...

Development

Prerequisites

• install opencv
• have a GPU
• install cuda
• python3

Run the following command to setup the project on your machine:

$ make setup

Labeling

To start labeling images run the following command. It is recommended to split the datasets into smaller chunks (~300-400 images each). This helps with your motivation labeling the dataset aswell as the performance of the labeling program.

$ make label TABLE=leonhart KIND=black-green DS=01

Sourcing data

1. Fetch a video file from youtube or wherever e.g. using youtube-dl
2. Split the video file into frames

$ youtube-dl "https://www.youtube.com/watch?v=aAXxONJDB0A"
# ... writes a file to ./path/to/video-file

# now extract every nth frame from the video
# writing it to the pictures dir: ./leonhart/black-green/83/pictures
$ make split TABLE=leonhart KIND=black-green DS=83 INPUT=./path/to/video-file

# now, proceed with labeling
$ make label TABLE=leonhart KIND=black-green DS=83

Training the Dataset

# transform source dataset into darknet-readable format
$ make export

$ make train-docker

# ...or train without docker
$ make train