This page summarizes my work developing a raindrop segmentation pipeline, from manual image labeling to training a custom U-Net model for automated detection. Some deliverables are currently restricted due to research confidentiality.
Formal Research Report
A detailed report on the development of a manual labeling tool with SAM2 integration, the construction of a mask dataset, and the first stages of automated model training for raindrop detection.
Research Presentation
A visual walkthrough of my project’s evolution, highlighting technical decisions, segmentation performance, and ongoing improvements like horizon-based data sampling.
These deliverables summarize my undergraduate research conducted under the guidance of Dr. Yan Tong at the University of South Carolina and supported by the Office of Naval Research. The formal research report outlines the complete development of a custom raindrop segmentation pipeline — beginning with a Python-based image labeling application enhanced by OpenCV and SAM2 model integration to generate precise segmentation masks.
The report documents the construction of a ground truth dataset from over 50 real-world images and describes the transition to automated segmentation using a U-Net-based model architecture. It includes implementation details, data preprocessing strategies, and refinements introduced based on evaluation feedback. Notably, the report discusses the use of horizon-line detection to support targeted data sampling — a strategy that improved model performance in distinguishing between similar visual features across sky and water surfaces.
The accompanying presentation slides visually trace this evolution using annotated architecture diagrams, mask overlay examples, training process screenshots, and performance benchmarking. It highlights technical challenges such as annotation efficiency, mask consistency, and robustness under varying lighting and scene conditions. Both deliverables emphasize the iterative development process and a forward-looking approach toward scalable, real-time segmentation pipelines for vision systems operating in dynamic environments.
Due to internal review protocols and data classification policies, these materials are currently designated as Controlled Unclassified Information (CUI). They will be made publicly available upon completion of the project and final approval.