Object Detection on Three Dimensional Reconstructions of Coral Reefs

Coral Reefs are structurally complex ecosystems that have been severely affected by natural and anthropogenic stressors. Consequently, there is a need for rapid and accurate ecological assessment of coral reefs, but current approaches entail time-consuming manual data acquisition and analysis. This thesis proposes an algorithmic approach to identify coral entities in 3D ecosystem maps as distinct 3D objects. Given 2D region proposals in an RGB image and an annotated 3D reconstructed surface mesh, our method generates a 3D region proposal for every 2D region proposal using the intrinsic and extrinsic camera parameters associated with the RGB images. The performance of the proposed method on coral reef survey images is compared against the performance of a state-of-the-art end-to-end learning-based approach called the Frustum PointNet and the results are tabulated and compared. The average precision values obtained using the proposed method and Frustum PointNet are observed to be dependent on the underlying coral class.