About
This project takes place in the context of a collaboration with the microbiologists with whom we are intending to develop a pipeline to detect and track bacteria attaching themselves on a glass substrate. In this experimental setup, a flow of water loaded with bio-luminescent bacteria is passing a glass substrate while being imaged with a microscope at 0.5Hz. Some bacteria attach to the surface, sometimes detaching after some time. The objective of the project is to build a detector and tracker to follow these events and provide statistics on the adherence properties of the bacteria in various conditions, as a function of their angles with respect to the flow in particular.
Procedure
We started with a set of 500 images and applied a basic green threshold to quickly create a preliminary dataset. With this dataset, we trained a small YOLOv5 model. We then used Model Assisted Labeling on Labelbox to further refine our model. We incorporated a Kalman filter and bounding box overlap matching step for tracking, including orientation. Finally, we saved all metadata to CSV files and used Pandas to build statistics.
Skills developped
✅ Object detection
✅ Pytorch
✅ Datasets creation, Data Analysis
✅ MAL
✅ Communication