SoccerNet Ball Action Spotting Challenge 2024 Baseline

This repo is a fork of the baseline for the SoccerNet Ball Action Spotting Challenge 2024 for the SoccerNet Ball Action Spotting Challenge 2024.

This repo stores the solution of SoccerNet Ball Action Spotting Challenge 2024 by Team Ai4sports. We tried some methods to improve the ability of classification.

Quick setup and start

Requirements

• Linux (tested on Ubuntu 20.04)
• NVIDIA GPU (pipeline tuned for RTX 3080)
• NVIDIA Drivers >= 520, CUDA >= 11.8
• Docker
• NVIDIA Container Toolkit

Run

Clone the repo and enter the folder.

bash git clone git@github.com:recokick/ball-action-spotting.git cd ball-action-spotting

Build a Docker image and run a container.

bash make

From now on, you should run all commands inside the docker container.

Download sampling_weights_001 and action_sampling_weights_002 from author's Google Drive and copy the files to the data directory so that the folder structure is as follows:

data ├── action │ ├── experiments │ │ └── action_sampling_weights_002 │ └── predictions │ └── action_sampling_weights_002 ├── ball_action │ ├── experiments │ │ └── sampling_weights_001 │ └── predictions │ └── sampling_weights_001 ├── readme_images └── soccernet └── spotting-ball-2024 └── england_efl

OR

Download the Ball Action Spotting 2023 dataset and Action Spotting 2023 dataset if you want to train the models from scratch. To get the password, you must fill NDA (link).

Now you can train models and use them to predict games. To reproduce the final solution, you can use the following commands (for the --experiment sampling_weights_001 parts of the steps you might want to change the constant soccernet_dir in src/ballaction/constants.py to `soccernetdir / "spotting-ball-2023"`):

```bash

Train and predict basic experiment on all folds

python scripts/ballaction/train.py --experiment samplingweights001 python scripts/ballaction/predict.py --experiment samplingweights001

Training on Action Spotting Challenge dataset

python scripts/action/train.py --experiment actionsamplingweights_002

Transfer learning

python scripts/ballaction/train.py --experiment balltuning001 python scripts/ballaction/predict.py --experiment balltuning001 python scripts/ballaction/evaluate.py --experiment balltuning001 python scripts/ballaction/predict.py --experiment balltuning001 --challenge python scripts/ballaction/ensemble.py --experiments balltuning_001 --challenge

To train models without the use of the test subset of the data in training use argument --folds train

Spotting results will be there

cd data/ballaction/predictions/balltuning001/challenge/ensemble/ zip resultsspotting.zip .///*/results_spotting.json ```