tennis-play-vs-nonplay

Tennis Play vs Non-Play Classifier

Model Description

This is a fine-tuned ResNet-18 model designed to classify frames from tennis broadcast videos into two categories:

• Play: Frames where a tennis rally is actively taking place.
• Non-Play: Frames containing crowd shots, player close-ups, commercials, or breaks.

This model was trained to help automate the creation of tennis match highlights by filtering out all non-action sequences.

Intended Use

• Action Filtering: Automatically filter long tennis match videos to keep only active play segments.
• Highlight Generation: Pre-processing step for automated sports highlight pipelines.

Training Data

The model was trained on a custom dataset extracted from grand slam matches (Wimbledon, US Open, Roland Garros). Frames were extracted and manually sorted into Play and Non-Play directories to create a robust ground-truth dataset.

Performance

• Base Model: ResNet-18 (pre-trained on ImageNet)
• Accuracy: ~94% on validation set
• Inference Speed: ~35 FPS on standard GPU

How to Use (PyTorch)

import torch
import torchvision.models as models
import torch.nn as nn

# 1. Load the ResNet-18 architecture
model = models.resnet18()
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, 2) # 2 classes: Play and Non-Play

# 2. Load the trained weights
state_dict = torch.load('play_vs_nonplay.pth', map_location='cpu')
model.load_state_dict(state_dict)
model.eval()

# 3. Model is now ready for inference!