Project Description
A smart bin powered by a cloud powered vision based garbage classification system
Device prototype
Vision sytem output
Architecture
System Architecture
AI Architecture
Role
Planned and executed the updation and upgradation of the in-house developed vision based Garbage Sorting System for alpha launch of the product.
Improved the data acquisition and annotation strategy; increasing training data from 4000 to 14000, maintaining 100% data integrity while merging the old and new data set by developing a custom web scrapper, devising the first data-class mapping diagram to assist annotators, using python and Amazon Mechanical Turk Updated system architecture; increasing accuracy from 70% to 83%, increasing predicted class labels from 3 to 15; by replacing original Mask-RCNN with YoloV4 network fine-tuned on custom data set
Outcome
Organized stress testing of the deployed model on Ali Baba; simulating 1500 concurrent users, identifying required hardware; using Locust.py
Tech stacks used
Machine Learning, Data Analisys, REST/RESTful APIs, AWS Glue, Flask, Python, Docker, Modular APIs, Azure, ETL, Pandas, Numpy, Git
