Final Project: Visual Similarity Between Movie Posters

Implementation of Classifying_Movie_Poster_Genres_with_Deep_Metric_Learning.pdf

---

Project Highlights

• leverage a base ResNet model combined with proxy anchor loss to learn an embedding space for movie posters.

Previous Work

Research draws upon previous work in convolutional neural networks, movie genre classification, and metric learning. The ResNet architecture, proxy anchor loss, and deep metric learning methods are some of the notable techniques utilized.

Technical Approach

The following key steps make up the project workflow:

• Data Collection and Preprocessing: Movie poster dataset collected from themoviedb.org and preprocessed for uniform size, normalization, and data augmentation.
• Network Architecture: Using ResNet50 as the backbone for our embedding model.
• Learning Method: Employing the Proxy Anchor Loss to train the ResNet50-based embedding model.
• Model Training: Utilizing PyTorch and PyTorch Lightning for training.
• Evaluation and Testing: Evaluating the model with Recall@K metric.
• Similar Movie Posters Retrieval: Using Approximate Nearest Neighbors (ANN) to retrieve similar movie posters.

Dataset and Implementation

Created a new movie dataset by leveraging themoviedb.org's API. This decision was because existing public movie datasets did not contain recent movie releases and had issues with acquiring the movie poster image files. The API provided data for 10,000 popular movies.

Getting started with development

Prereqs:

• Python 3.10+

Check if you have Python 3.10 installed

python3.10 --version

If you get python3.10: command not found

try: python3 --version

If you get Python 3.X.Y where X is at least 10 then update the last line in the Pipfile to your output and continue to next section

ex:

main@LAPTOP:~/cosc525_final_project$ python3 --version
Python 3.10.6

# Pipfile
[requires]
python_version = "3.10"

Follow: Installing Python 3.X, but install the Python 3.10.6 binary

Should be this page - https://www.python.org/downloads/release/python-3106/

Python Virtual Env

Create the virtual environment & install the requirements:

# this installs numpy automatically
pipenv install --dev

Enter the virtual environment:

pipenv shell

Launch Jupyter notebook environment (Mostly for local development):

# This launches the notebook env & should open your browser
jupyter notebook

License

Visual Similarity Between Movie Posters is available under the MIT license. See the LICENSE file for more info.

Author

Developed by Edgar Trujillo