NFL Challenge: Detect helmets in American Football

This projects contains a pretrained EfficientDet that detects helmets in football images.

References

EfficientDet from PyTorch repo.

Dataset

The data were provided by th challenge orginisers. The data can be downloaded from the NFL Impact Detection. The script to download the dataset is in scripts/download_data.

To generate the frames in the right place and format, you can run python src.prepare_data.py, that will take the raw videos from the ../../data/train folder and create images. Feep in mind, it requires disk space of 65 GB for the full dataset.

Prepare environment

Clone this repo
Install anaconda
Create conda environment
Run pip install -r requirements.txt

Make folds

All folds are in src/folds directory, make_folds.py creates folds

Inference

For MP4 videos, you have to go through every frame and run the prediction. Use src/prepare_data.py to generate frames from videos.

For images, run: python inference.py runs inference for helmets detection

Reproduce Training

Generate training frames.

One class helmet detection

Change path to data in src.effdet_train.py

First, train 1-class helmet detection: python src.effdet_train.py

Two class helmet detection

Change path to wieghts in src.effdet_train_2classes.py

Fine-tune detection for 2 classes: python src.effdet_train_2classes.py

Second-level 1D CNN classifier

Use helmet predictions/ground truch for second-level 1D CNN classifier. Generate predictions tensor and train 1D CNN: python cnn1d.train_trajectories.py

Second-level 3D CNN classifier

Use helmet predictions/ground truch for second-level 3D CNN classifier. Generate predictions tensor and train 1D CNN: python cnn3d.train_video_classifier_pytorch.py

tatigabru/nfl_impact_detection