ECSE 415 - Introduction To Computer Vision Final Project

Nathan Clairmonte, Ali Shobeiri, Tiffany Wang, John Wu, Frank Ye.

The following repository holds the source code for the ECSE 415 - Introduction To Computer Vision course final project. The final project involves completing both the classification and localization on the MIO-TCD dataset with specific method restrictions. The exact requirements can be found in the attached PDF file.

Prerequisites

YOLO v3 (You Only Look Once)

A custom trained model of the YOLO v3 deep learning implementation is utilized for both localization and classification. The model was trained by using the provided ground truth data and following the instructions in the Darknet github repository.

Download the latest trained weights and place it in the src/yolo directory or download directly from a linux terminal.

NOTE: the trained weights are deleted.

MIO-TCD Data

Both the localization and classification datasets will be used in the challenge. Download them directory from the website or through a linux terminal.

wget http://podoce.dinf.usherbrooke.ca/static/dataset/MIO-TCD-Classification.tar
wget http://podoce.dinf.usherbrooke.ca/static/dataset/MIO-TCD-Localization.tar

In addition to the provided datasets, an additional dataset was generated using the ground truth predictions for the localization dataset and can be downloaded here. The dataset can also be downloaded directly through a linex terminal.

wget https://415.blob.core.windows.net/data/localizations_cropped.zip

Python Packages

The necessary python packages can be installed by running the requirements.txt file using pip.

pip install -r requirements.txt

Source Code

All source code for the project can be found in the src folder. The main root folder contains three main scripts for outputting the necessary results to complete the goal of the challenge.

• classifier.py: Runs the trained classifier model k-folds through the classification dataset. This code works with both SVM and Logistic Regression models. Use the following command to run the code:

python classifier.py -d <data_path> -o <output_directory> -c <classfiier / svm or logreg> -k <number of k-folds>

All the arguments are mandatory except for the number of kfolds.

• localizer.py: Runs the trained yolo model through a test localization set while also feeding the localization outputs to the trained SVM classifier. This script outputs results for localization (using YOLO) in addition to classification (using both YOLO and SVM).

In addition to the main scripts are a set of directories used for various tasks:

• classifier: Contains all source code used for preprocessing, training, and testing the SVM and logistic regression classifiers.
• yolo: Contains all source code used for preprocessing, training, and testing the custom YOLO model.
• results: Output results used for the report, including screenshots of the performance of each model and a utility script used for additional plotting and organization of the results.