All22 Computer Vision

Computer Vision tags on all 22 football film

Purpose

Create models off to predict coverages, blitz, qb spies, offensive personnel, zone/man, etc. utilizing native film. The goal is to help to reduce time tagging videos by predicting coverages with confidence

Inspiration: https://www.opensourcefootball.com/posts/2021-05-31-computer-vision-in-r-using-torch/

Process

1. Find All 22 NCAAF Football Film
2. Slice Film by ESPN Game ID (to join for future tagging with @cfbfastR data)
3. Add Training Data w/ manual tags (train_labels)
4. Build Models using CNN model architecture and with base architecture of ImageNet (VGG16)
5. Test w/ separate tags (test_labels)

Games currently sampling plays to train model (list will continue to grow)

(At Least 10 plays of each game are represented in the dataset)

• 2022 Alabama vs Georgia National Championship Game (https://www.espn.com/college-football/game/_/gameId/401331242)
• 2020 Clemson vs LSU National Championship Game (https://www.espn.com/college-football/game/_/gameId/401135295)
• 2020 Baylor vs Georgia Sugar Bowl (https://www.espn.com/college-football/game/_/gameId/401135289)
• 2019 Nevada vs Hawaii (https://www.espn.com/college-football/game/_/gameId/401117515)
• 2019 Oregon vs University of Washington (https://www.espn.com/college-football/game/_/gameId/401114196)
• 2017 Ohio State vs Indiana Season Opener (https://www.espn.com/college-football/game/_/gameId/400935229)
• 2015 Baylor vs Oklahoma State (https://www.espn.com/college-football/game/_/gameId/400763462)
• 2014 TCU vs Kansas (https://www.espn.com/college-football/game/_/gameId/400547892)

*Training set: 130 plays *Test set: 25 plays

Defensive Coverage Model Architecture

Utilizing VGG-16 Image Net as the base model architecture: https://www.geeksforgeeks.org/vgg-16-cnn-model/

Then add additional layers on top of this model to build the learning specific to our "manual charted" data:

Model Performance

54% on test set so far (Majority of coverages are Cover 2)

Helper Model (ImageNet + Custom Architecture): Training Accuracy: 92% Custom Architecture: Training Accuracy: ~46% (Need improvement)

To - Do

• Fix Input shape for testing off of custom architecture
• Find more instances of Cover 3/4
• Add Blitz Probability

Sample Output