Vedanth Kallakuri
Research Mentor(s): Katherine Skinner
Mentor Department: Robotics
Authors: Vedanth Kallakuri, Cale Colony, Jingyu Song, Katherine Skinner
Session: Session 4 (1:00pm – 1:50pm)
Presentation Type: Poster 53
Abstract
Understanding properties of fish populations plays a vital role in society, contributing to economic growth, nutritional needs, and sustainability efforts. For example, analyzing fish length, width, and mass is integral to maintaining a healthy stock in the aquaculture industry. Meanwhile, counting wild fish populations fuels marine conservation and coastal reef restoration efforts amidst vast habitat destruction caused by overfishing and climate change. This project proposes developing innovative machine learning and artificial intelligence technologies to improve fish detection and classification in underwater environments, which would provide critical information for marine science and industry applications. Traditional efforts, such as manual fish measurements, are labor intensive and inefficient. Alternatively, deep learning-based segmentation models automatically classify pixels of each fish into meaningful regions, offering a promising solution. Most of these models, however, are fully supervised, requiring costly and time-consuming per-pixel segmentation labels, where every pixel of each fish has to be annotated as training data. To address this, we propose a weakly supervised segmentation model that only requires point labeled data to produce accurate segmentation masks of fish in an underwater scene. Point labeling only requires annotating the center of each fish, reducing the burden of per-pixel labeling of each fish. The developed model will be tested on several datasets from real environments, and will be evaluated against the produced segmentation masks of existing baseline models. The mIOU (mean intersection over union) metric will be used to quantify and compare segmentation mask accuracy.
