Machine Learning Sports Classification

A comprehensive machine learning project for classifying 100 sports categories from images using both deep learning (CNNs) and traditional machine learning approaches.

---

Overview

This project implements and compares multiple approaches for sports image classification:

• Deep Learning (CNNs): ResNet, EfficientNet, Vision Transformer
• Traditional ML: KNN, Random Forest, SVM
• Feature Extraction: Color Histograms, HOG, LBP
• Dataset Analysis & Augmentation utilities

Dataset size: 14,899 images across 100 sports categories.

---

Dataset

Dataset Statistics

• Training: 11,958 images
• Validation: 1,405 images
• Test: 1,536 images
• Total: 14,899 images

Directory Structure

MachineLearning-SportsClassification/
├── train/
│   ├── air hockey/
│   ├── archery/
│   └── ... (100 classes)
├── valid/
│   └── ... (100 classes)
└── test/
    └── ... (100 classes)

---

Quick Start

1. Environment Setup

Important: Use Python 3.10–3.12 (Python 3.13 is not supported by some ML libraries).

git clone https://github.com/asyau/MachineLearning-SportsClassification.git
cd MachineLearning-SportsClassification

python -m venv venv
source venv/bin/activate  # macOS / Linux
# venv\Scripts\activate  # Windows

pip install --upgrade pip
pip install -r requirements.txt

---

2. Dataset Preparation

Ensure your dataset is placed in the project root:

train/
valid/
test/

---

Dataset Analysis & Preparation (Optional)

Analyze Class Distribution

python analyze_data_distribution.py

Outputs:

• Class distribution plots
• Imbalance analysis
• CSV summaries (saved to class-dist-analysis/)

---

Get Class Sizes

python get_train_class_sizes.py

Outputs:

• train_class_sizes.txt

---

Augment Small Classes

Augments classes with fewer than 110 images.

python augment_small_classes.py

---

Create Train/Validation/Test Split

python create_train_val_test_split.py

Edit the script to configure:

• Input directory
• Output directory
• Split ratios (default: 80 / 10 / 10)

---

Deep Learning (CNN) – Recommended

Training

python train.py

or

cd cnn
python train.py

Configuration (cnn/config.py)

MODEL_NAME = 'efficientnet_b4'  # resnet50, efficientnet_b0-b4, vit_b_16
BATCH_SIZE = 24
NUM_EPOCHS = 15
LEARNING_RATE = 0.0005

USE_CLASS_WEIGHTS = True
USE_MIXED_PRECISION = True
EARLY_STOPPING = True
EARLY_STOPPING_PATIENCE = 5

---

Evaluation

python evaluate.py

Supported metrics:

• Top-1 / Top-3 / Top-5 Accuracy
• Precision, Recall, F1 (Macro & Weighted)
• Confusion Matrix
• Per-class metrics

---

Inference

Single image:

python inference.py --image path/to/image.jpg

Directory:

python inference.py --dir path/to/images/

Top-K predictions:

python inference.py --image path/to/image.jpg --top_k 10

Custom checkpoint:

python inference.py --image path/to/image.jpg --checkpoint checkpoints/best_model.pt

---

TensorBoard

tensorboard --logdir logs/

Open: http://localhost:6006

---

Traditional Machine Learning Approaches

K-Nearest Neighbors (KNN)

python knn_color_histogram_only.py
python knn_hog_only.py
python knn_lbp_only.py

---

Random Forest (RF)

python rf_color_histogram_only.py
python rf_hog_only.py
python rf_lbp_only.py

---

Support Vector Machine (SVM)

Edit SVM_hog_color_lbp.py:

BASE_DIR = "."
SELECTED_FEATURE = 'color'  # 'color', 'hog', 'lbp'
USE_PCA = True

Run:

python SVM_hog_color_lbp.py

Outputs saved to svm_output/.

---

Combined Features

python concatinated.py

Uses Color + HOG + LBP together.

---

Project Structure

MachineLearning-SportsClassification/
├── cnn/
│   ├── config.py
│   ├── dataset.py
│   ├── model.py
│   ├── train.py
│   ├── evaluate.py
│   ├── inference.py
│   └── utils.py
│
├── train.py
├── evaluate.py
├── inference.py
│
├── analyze_data_distribution.py
├── augment_small_classes.py
├── create_train_val_test_split.py
├── get_train_class_sizes.py
│
├── knn_*.py
├── rf_*.py
├── SVM_hog_color_lbp.py
├── concatinated.py
│
├── train/
├── valid/
├── test/
├── checkpoints/
├── logs/
└── class-dist-analysis/

---

Requirements (Core)

Key dependencies (see requirements.txt for full list):

• torch
• torchvision
• tensorflow
• scikit-learn
• opencv-python
• numpy
• pandas
• matplotlib
• seaborn
• tensorboard

---

Troubleshooting

CUDA Out of Memory

• Reduce BATCH_SIZE
• Reduce IMAGE_SIZE
• Enable mixed precision

Slow Training

• Use smaller model (efficientnet_b0)
• Enable mixed precision
• Increase NUM_WORKERS

Dataset Not Found

• Ensure train/, valid/, test/ exist in project root
• Check BASE_DIR in scripts

---

Notes

• CNN models give significantly better accuracy than traditional ML
• Traditional ML is useful for baseline comparisons
• Class imbalance handling is strongly recommended

---