Deep Learning Models

Game Status Detection

In sports analytics, it's often important to catch the events. In volleyball, just like any other sport, there are a lot of gaps between the events. We need to catch the volleyball events when the game is in play, so to do this, we need to formulate it.

In this repo, we formulated it this way:

we sliced each video into 1-second clips and extracted each clip into its frames, which would become 30 frames each. Then we input them into a fine-tuned version of VideoMAE model on a custom dataset, and it outputs 1 label denoting service, in-play and no-play.

The label service is the state of the game that the player tosses the ball, and hits it to start the rally.

The label in-play denotes the game states that the players are playing, the event is in progress, and it hasn't been stopped.

The label no-play denotes the game states that there is no event or the players are just warming up and not playing for the score.

Model Architecture:

videomae architecture

Dataset:

train:

no-play	play	service
7136	4886	2447

test:

no-play	play	service
794	544	273

The training results

Confusion matrix:

The evaluation results:

eval_accuracy	eval_precision	eval_recall	eval_f1	eval_loss
0.9906890130353817	0.9911190984494304	0.9883168656922177	0.9896984571368996	2.413252830505371

Ball Segmentation

Yolov8 is the state-of-the-art architecture in the object detection realm and it's developed by ultralytics group and is used for all object detection/segmentation tasks in this repo.

Datasets:

Dataset1:

volleyVision ball detection