UMARV-CV-ScenePerception

A research and development platform for the University of Michigan's Autonomous Robotic Vehicle (UMARV) Computer Vision team to solve the understanding of a road scene.

Models vs Algorithms

The models folder hosts all of our machine learning solutions, while the algorithms folder hosts our hard coded solutions. Each model/algorithm is seperated into its own folder and has its own unique ID.

Scripts

The src/scripts/ folder hosts our scripts which provide varrying functionalities from model/algorithm initialization, performance comparison, and dataset generation. To run them, right click on the script and select "Run Python File in Terminal".

How To Interact With This Repository

Video Tutorial Coming Soon!!

1. Have git installed on your computer.
   • git installation guide
   • git introduction
2. Have Python installed on your computer.
   • Python Installation Guide
3. Request access to the ScenePerception GitHub repository from a team lead.
   • You must accept the invitation to the GitHub repository.
4. Setup the repository on your local machine.
   • On your Desktop, right click and select 'Open In Terminal'.
   • mkdir UMARV
   • cd UMARV
   • mkdir ScenePerception
   • cd ScenePerception
   • git clone https://github.com/umigv/UMARV-CV-ScenePerception.git
   • cd UMARV-CV-ScenePerception
   • IMPORTANT: Replace your branch name in the end of the next 2 commands.
     • your_branch_name = "user/{your_name_with_no_spaces}"
     • Ex: Branch name for Awrod Haghi-Tabrizi = user/AwrodHaghiTabrizi
   • git checkout -b {your_branch_name}
   • git push -u origin {your_branch_name}
5. Open the project in VSCode.
   • Open VSCode.
   • Click File > Open Folder.
   • Open the UMARV-CV-ScenePerception folder.
     • Common mistake: Opening the UMARV folder or the ScenePerception folder.
     • IMPORTANT: Keep your working directory as UMARV-CV-ScenePerception when running scripts and notebooks.
6. Before starting development, install the following Python libraries on your machine / virtual environment either manually or with pip install -r requirements.txt.

For New Members

Before creating your first model its suggested that you create one copy of the model_template and use one of the notebooks to train a small model. Here are the instructions on how to do this:

1. Follow this tutorial for creating a DropBox developer app: Tutorial
2. Run src/scripts/create_copy_of_model.py and when asked for a unique identifier, input template.
3. Notice the new model folder that was created in /models/ and briefly look through the architecture.py, dataset.py, and methods.py. An explanation of each of these files can be found here.
4. Select a notebook you want to work with in the notebooks directory. It's recommended that you try the colab_env first as it is the most common one we will use. Visit this link to understand how to use it.
5. Verify your notebook is running on a T4 GPU. This can be found in the top right of the screen just left of the Gemini AI logo.
6. Run each cell of the notebook and train a template model. Keep the epochs under 10 at first so that the training doesn't take too long. Then make sure you were able to visualize the model output in the notebook (it is expected to have very bad performance at first).

Repository Rules

• Full freedom to create/delete/edit code in your model/algorithm folder.
• Dont change any code in:
  • model/algorithm folders that dont belong to you (you can tell by the author name in the content/info.json or just by the model id itself).
  • src/scripts/ (unless making global updates).
  • model_template/algorithm_tempalte (unless making global updates).
• Work in your own branch. Pull before every work session. Push after every work session.

Environments

This repository allows development flexability to work in multiple environments, including: - Windows - Mac - Google Colab - Working with Google Colab - LambdaLabs - Working with LambdaLabs - Jetson (coming soon)

Developing Models

1. Navigate to src/scripts.
2. Right click on either create_model.py or create_copy_of_model.py
   • create_model.py creates a new model from the template
   • create_copy_of_model.py creates a copy of a model using its model id
3. Click "Run Python File in Terminal" OR run python3 src/scripts/create_copy_of_model.py in your terminal
4. Answer the prompts in the terminal. When it asks for the model ID, include only the unique identifier of the model you want, do not include model_.
5. Go through Working With Models

Machine Learning Model Leaderboard

This leaderboard showcases the top performing segmentation models developed by UMARV members based on the average accuracy of the model during testing. Once you have developed a PyTorch model that has higher accuracy, you are free to add your model to the leaderboard by editing the main branch README.md. Still keep the model folder in your personal branch only.

#	Name	Accuracy	Other Metrics	Creators	Git Branch Name
1	32mw3qk4	0.2390	Mean IoU: 0.1083; Mean Dice Coeffecient: 0.1872	Awrod	Main

Developing Algorithms

1. Navigate to src/scripts
2. Right click on either "create_new_algorithm.py" or "create_copy_of_algorithm.py"
3. Click "Run Python File in Termainl"
4. Answer the prompts in the terminal
5. Go through Working With Algorithms