CanSat Picture-Segmentation

Object-Segmentation model created for Project Trailblazer’s project in 2019/2020 CanSat competition. Program divides pictures taken by our probe into categories such as forest, grass, sand.

In this project we used Python and it's libraries, eg.: NumPy, PIL, PyTorch, TensorFlow

Folder structure

We created two different models, each using different machine learning solutions.

Both projects contain

• Data

Folder containing photos used for training our models.

• DataPreparation.ipynb

We needed to create a dataset containing substantial amount of photos. Therefore, we created a simple program that loads a large picture, resizes it to a multiplicity of 224 and cuts it into smaller pictures (224x224px)

• App.ipynb

Program that loads photos made by probe and pre-trained weights to generate final output of this branch of project.

• Test.png

Photo to test both models.

Project A

• Model.ipynb

Program that loads photos from data folder and uses RandomForestClassifier to generate weights for the machine-learning model.

Project B

• ResNet18.ipynb

Program that loads photos form data folder and pre-trained ResNet18 model. We have had the dense layer cut and replaced it with our own one. Then it saves generated weights

Example

Photo made by CanSat's camera
Output of the program

Author

Adam Kaniasty as a member of Project Trailblazer Team.