Inspiration
I wanted to tackle the growing problem of waste mismanagement and recycling inefficiency. Seeing how much time and effort is wasted sorting trash manually, I thought: why not use computer vision to make recycling smarter and faster?
What it does
Grevion uses computer vision to detect and classify different types of waste—plastic, metal, cardboard, glass, and general trash—in real-time. It provides immediate feedback on bin fullness and helps automate the sorting process to improve recycling efficiency.
How we built it
I combined YOLOv5 for object detection with a MobileNetV2 classifier as a verification layer. My dataset includes thousands of labeled images across multiple waste categories. I used OpenCV for webcam integration, real-time tracking, and visualization. The system is designed to be robust under different lighting conditions, rotations, and partial occlusions.
Challenges we ran into
Handling lighting variations, occlusion, and rotation of objects in real-time video. Aligning YOLO predictions with the MobileNetV2 classifier to improve accuracy. Managing a large dataset and ensuring labels were consistent without manually annotating every image.
Accomplishments that we're proud of
Built a fully functional real-time smart waste detection prototype. Implemented a double-check classification system that significantly improves accuracy. Created a pipeline that can easily scale to new categories of waste without retraining everything from scratch.
What we learned
The importance of robust dataset preparation for real-world object detection. How to integrate multiple computer vision models (YOLO + MobileNetV2) effectively in a real-time system. Techniques for handling variable lighting and occlusion in computer vision projects.
What's next for Grevion
I plan to optimize the model for edge devices so it can run on low-power hardware. I also want to expand the dataset to include more waste types and integrate smart bin alerts to inform users when a bin is nearly full, helping municipalities manage waste more efficiently.
Log in or sign up for Devpost to join the conversation.