This repository is related to the project for the Machine Learning course, which was held at the Department of Mathematics and Computer Science of the University of Catania by prof. Giovanni Maria Farinella.
The project development team is composed of Lorenzo La Rocca, Raffaele Terracino, and Edoardo Tantari.
The goal of the project is to build and compare different object detection models capable of detecting and classifying road signs. For this project, a subset of 19 classes was selected from all possible European road signs.
This subset includes the following road signs:
- mandatory left turn sign
- mandatory right turn sign
- yield sign
- no entry sign
- stop sign
- green traffic light
- red traffic light
- speed limit signs from 20 to 120 km/h
Among the state-of-the-art one-stage architectures, YOLOV12 was selected. For the two-stage approach, Faster_RCNN_ResNet50 was used.
demo/– contains the streamlit web app for running inference on images and videosdocs/– contains the technical report describing data collection, models, experiments, source code and demomodels/– contains metrics and checkpoints of trained modelsscripts/– contains scripts for dataset pre-processingsrc/– contains the source code for training and evaluating the modelsyolo/- contains a notebook for training, evaluation and inference on YOLO v12 using the Ultralytics libraryFaster_RCNN/- contains a notebook and scripts for training, evaluation and inference on Faster R-CNN, written using Pytorch
requirements.txt– list of python dependencies to run scripts, notebooks and the demo
After installing the dependencies listed in requirements.txt, move to demo/ and run the following command:
streamlit run main.py
The first time you launch the web app, it will be slower than subsequent launches due to the download of the trained 315 MB FasterRCNN model from HuggingFace.