PlayerDirection SqueezeNet Model

Overview

This model is trained for ice hockey player orientation detection, classifying cropped player images into one of eight orientations: Top, Top-Right, Right, Bottom-Right, Bottom, Bottom-Left, Left, and Top-Left. It is based on the SqueezeNet architecture and achieves an F1 score of 75%.

Model Details

• Architecture: SqueezeNet (modified for 8-class classification).
• Training Configuration:
  • Learning rate: 1e-4
  • Batch size: 24
  • Epochs: 300
  • Weight decay: 1e-4
  • Dropout: 0.3
  • Early stopping: patience = 50
  • Augmentations: Color jitter (no rotation)
• Performance:
  • Accuracy: ~75%
  • F1 Score: ~75%

Usage

1. Extract frames from a video using OpenCV.
2. Detect player bounding boxes with a YOLO model.
3. Crop player images, resize them to 224x224, and preprocess with the given PyTorch transformations:
   • Resize to (224, 224)
   • Normalize with mean=[0.485, 0.456, 0.406] and std=[0.229, 0.224, 0.225].
4. Classify the direction of each cropped player image using the SqueezeNet model:

   with torch.no_grad():
       output = model(image_tensor)
       direction_class = torch.argmax(output, dim=1).item()