Advanced Machine Learning Final Project a.y 2021-2022
La Sapienza University of Rome
MSC IN DATA SCIENCE
Daniel Jimenez, Juan Mata Naranjo, Alessandro Quattrociocchi, Tansel Simsek
Current models to count objects on imagesare often based on pre-trained models and density estima-tion, however they are still not close to optimal. Our pro-posal we first aim at reducing the gap between trainingand test error by introducing regularization techniquessuch Batch Normalization, Dropout and Data Augmenta-tion. In addition, to enhance the behaviour of the model,we proposed to use diverse ImageNet pre-trained mod-els (i.e. VGG16) as an alternative for ResNet50. As afinal novelty, we implemented an ensemble method bycombining ResNet with YOLO to produce a model thatoutperforms the current state-of-the-art work
└── Final_Project
├── README.md
├── data
│ ├── ImageClasses_FSC147.txt
│ ├── Train_Test_Val_FSC_147.json
│ ├── annotation_FSC147_384.json
│ └── pretrainedModels
│ └── FamNet_Save1.pth
├── funcs.py
├── interpretability_captum.py
├── main.ipynb
├── model.py
├── model_explainability.py
└── utils_ltce.py