This is the first stage of your robo-goose. Keep taking pictures of geese to see it's other stages!
This is the welcome/start page

Goosemon Go 📱✨

What It Does

Goosemon Go is a fun and easy-to-use app where you:

Own and care for a virtual robot goose.
Snap pictures of geese in your local neighborhood to progress and unlock cosmetic upgrades for your robot goose.
Collect and mark geese locations to help track population and migration patterns for potential research purposes.

Designed prototypes using Figma and hand-drew Goosemon Go assets (e.g., robot goose sprites) with Procreate.
Built using React and Vite with TypeScript for seamless UI interactions.
Styled with TailwindCSS and custom CSS animations for a modern, responsive design.
Enabled real-time communication with WebSockets (Socket.io).

Developed using Python and Flask, hosted on an AWS EC2 instance.
Flask backend integrated PyTorch to preprocess and classify images of geese.
Combined the Roboflow API for pretrained geese classification with a custom Convolutional Neural Network (CNN) for improved accuracy.
Implemented game logic for leveling up, earning experience points, and unlocking rewards.

Used a pretrained Roboflow model API to classify images of Canadian geese and baby geese.
Built a custom CNN trained on:
- ~1.4k 128x128 geese images as positive examples.
- ~1.1k 128x128 images of desks as negative examples.
Used OpenCV to dynamically capture and save negative examples to train a diverse dataset.
Preprocessed data with PyTorch by resizing, grayscaling, and normalizing images.
Achieved ~93% accuracy with the CNN, trained on 2,000+ labeled images and tested on ~800 labeled images.
Combined the custom CNN and Roboflow model to reduce false positives and false negatives.

False Classifications: Roboflow’s model struggled with false positives due to a lack of negative examples. We trained a custom CNN and used OpenCV to gather more negative examples to address this.
Network Restrictions: Devices couldn’t communicate due to campus network restrictions (Eduroam), requiring alternative solutions.
AWS Storage Issues: The AWS EC2 instance ran out of space due to a small /tmp buffer. We resolved this by redirecting temporary files to the user directory.
Data Quality: Online images were often blurry or too small, requiring significant curation for training datasets.
Flask Errors: Flask’s development server encountered terminal attribute handling issues, causing runtime errors.

Successfully built and trained a custom CNN with ~93% accuracy.
Learned and implemented Flask for the first time.
Integrated WebSocket communication into a Flask API.
Designed and deployed a polished React Native web app with a visually appealing UI.

How to preprocess large image datasets using PyTorch.
Building efficient and robust apps with React Native.
The challenges and nuances of deploying apps, especially React Native on AWS.
Designing and implementing Flask APIs.
Training binary image classification models with PyTorch and optimizing data preprocessing for better accuracy and faster training.