Sport Detection And Video Processing

• This porject aimes to make a deep learning model in order to extract items from a sport video in order to detect which sport is in the video .

• In this Model we will use a dataset of images classed by names, for example :

  {
      'Basketball': [
          image_1, 
          image_2, 
          image_3,
          ...
      ],
      'Football'  : [...],
      'Hockey' : [...],
      ...
  }

• We exctract the images in order to train the CNN with PyTorch .

• We will save the model and use it in video processing and in order to detect the sport playing in the video .

Python Libraries

• We used Pytorch for model creation, and tensor creations :

  

• We used sklearn for metrics and performance :

  

• We used numpy for ndarray manupilation befor transforme it to torch tensors :

  

Additional Details Of The CNN Model

• SGD (Stochastic Gradient Descent) Optimizer:

The SGD optimizer is a widely used optimization algorithm for training machine learning models. The update step for each parameter is computed using the following formula:

$\theta_{new} = \theta_{old}-(\alpha \times \nabla\mathcal{L(x, \mathscr{C}) - (\mathrm{momentum } \times \mathrm{previousUpdate} ) - (\mathrm{weightDecay} * \theta_{old})})$

with :

• $\theta_{new}$ : is the updated value of the parameter.

• $\theta_{old}$ : is the current value of the parameter.

• $\alpha$ : is the learning rate hyperparameter, controlling the step size during updates.

• $\nabla\mathcal{L(x, \mathscr{C})}$ : is the gradient of the loss function with respect to the parameter.

• $\mathrm{momentum }$ : is the momentum hyperparameter, which smooths the update process and helps to accelerate convergence in certain cases.

• $\mathrm{previousUpdate}$ : is the accumulated previous update for the parameter.

• $\mathrm{weightDecay}$ : is the weight decay hyperparameter, which applies L2 regularization to the parameter to prevent overfitting.

in Python and Pytorch :

optimizer = optim.SGD(model.fc.parameters(), lr=0.001, momentum=0.9, weight_decay=1e-4)

• CrossEntropyLoss:

CrossEntropyLoss is a common loss function used for classification tasks. The formula for the CrossEntropyLoss is as follows:

$\mathcal{L}(x, \mathscr{C}) = \frac{-\log(e^{x_c})}{\sum_{i=1}^{n}e^{x_i}}$

with :

• $x$ is the output (logits) from the last layer of the neural network for a particular input.

• $\mathscr{C}$ is the true class label of the input.

• $x_{c}$ is the input for the class $\mathscr{C}$.