3D Reconstruction, Segmentation and Classification of Corals from Aerial Images Final Report: CS 231A

Coral reefs cover an area of over 280, 000km and support thousands of species in what many describe as the “rain forests of the seas”. Coral reefs face numerous threats, with current estimates suggesting that thirty percent of corals are in critical condition and may die within 10 to 20 years [cite]. One resourceful way to monitor the health of a coral reef is to periodically take aerial images of the reef from a UAV, use these images to construct a 3D model of the reef and monitor its evolution in time. Here we present a methodology for the 3D reconstruction, segmentation and classification of corals form aerial images taken by an UAV. The proposed methodology consists of building 3D point cloud from the images based on “structure from motion” approach, and then building the 3D surface to fit the point cloud. We segment the 3D surface based on the discrete mean curvature of the surface, cluster the coral containing segment into groups of individual corals using mean-shift algorithm and finally classify the corals into two classes of coral species namely, Branching and Porites using SVM. With the help of experts in the field of marine biology, we verify that clusters generated by our algorithm are indeed clusters of corals with high precision and recall. Using their labeling of our dataset, we achieve 76.6% success in the classification process. When compared to standard 2D techniques, we find that both the segmentation and classification processes benefit from using the 3D information.