This repository provides an implementation for a vehicle detection algorithm based on classical computer vision features (HOG, SIFT) and machine learning classifiers like support vector machines or gradient boosting classifiers. To install the required packages, run:
pip install -r requirements.txt
This code relies on the dataset shared for the 2023 Visual Computing Kaggle Challenge. train and test folders are supposed to be nested in the data folder.
The report for this project is available here
To reproduce my final submission, run:
python3 inference.py --test_dir data/test --clf_path models/gradient_boosting.pkl --scaler_path models/scaler.pkl --submission
This relies on a Gradient Boosting classifier trained on 64x64 patches extracted from the training dataset, and applied in a sliding window manner on test images.
To train a classifier, run:
python3 train.py --model gradient_boosting --bow --spatial --hist
Possible models are linear_svm, svm or gradient_boosting.
bag_of_words.py: Creates the vocabulary for Bag-of-SIFT features.dataset.py: Defines a class that simplifies the creation balanced dataset of patches either containing (postive samples) or not containing (negative samples) any vehicles and the extraction of features.inference.py: Defines the inference pipeline that applies the trained classifiers on a sliding window manner on test images.train.py: Launches training runs of models