This repository contains the final project of the Image Processing for Computer Vision course (AY 20/21) at the University of Naples Federico II.
The aim of the project is to perform Instance Segmentation on a subset of the Cityscapes dataset.
The project, developed in team of 4, is based on the Mask R-CNN architecture proposed by Facebook. To be more specific, we used a version of the network implemented by Matterport, pretrained on the COCO dataset.
- utils-cityscapes.ipynb contains some classes needed to adapt the network to Cityscapes dataset.
- inspect-dataset-cityscapes.ipynb shows an Explorative Data Analysis conducted at a preliminary stage.
- hp-tuning-cityscapes.ipynb shows the hyperparameter tuning stage, which has been performed using KerasTuner.
- train-cityscapes.ipynb shows model's training, performed with the best hyperparameters configurations found in the previous stage. For this purpose we exploited GPUs provided by Google Colab. Due to its high complexity and limited resources available, the model has been trained for a limited number of epochs.
- evaluation-cityscapes.ipynb shows the evaluation of the model performed on the test set. The best model reached an Average Precision (AP) of 0.623 with an Intersection over Union (IoU) threshold at 0.5.
- inference-cityscapes.ipynb shows how to use the model to make inference on test images.