An Investigation on the Required Perception for 3D Reconstruction of Branching Coral

Figure 1: The potential effects of bleaching on coral reefs. Image courtesy of G. Roff. Coral reefs are biologically-complex ecosystems that support a wide variety of marine organisms. These are fragile communities under enormous threat from natural and human-based influences. Sea surface temperatures have increased over the past few decades, resulting in widespread coral bleaching at an ever-increasing rate. The widespread mortality of corals following mass bleaching events reduces the structural complexity of reefs – eliminating the three dimensional habitat, which is critical to maintaining diversity and population of coral reef fish communities (Fig. 1). Despite the importance of coral reef ecosystems, spatial and temporal dynamics of coral bleaching events are poorly understood. To date, surveys of coral bleaching have been conducted using either human divers or remote sensing (satellite imagery), but these approaches are fundamentally limited in scale. Satellites are able to cover large spatial scales (10s of km2), and approximate percentage of coral cover, but are unable to resolve fine-scale ecological changes (cm to meters) of individual corals. In situ measurements by divers can provide this data, but at limited spatial scales (< 1 km2). Robotics provides a novel solution to the fundamental problem of large-scale area coverage, and represents a viable approach to quantifying the extent of fine-scale coral bleaching and reef structural complexity over large areas (> 100 km2). However, the fundamental question arises in how to plan and execute a sampling path for a robot to collect data in such a dynamic and structurally complex environment. In this paper, we provide an initial investigation into the perception required to create accurate 3-D reconstructions, based on visual imagery, for estimating physical parameters, e.g., surface area and volume, of complex reef features. Combining this information with the kinematic and dynamic constraints of the survey vehicle, we show how to compute a path for an underwater vehicle that acquires the images necessary to produce the best possible 3D reconstruction. We focus this paper on the analysis of a set of underwater images of a piece of staghorn coral, Acropora cervicornis. This dataset consists of 81 high resolution images of a piece of white Staghorn coral sitting on the floor of a pool, e.g., Fig 2(a). This type of coral is a branching, stony coral with cylindrical branches ranging from a few centimetres to over two metres in length and height. This is one of the three most important Caribbean corals, in terms of its contribution to reef growth and fishery habitat. Additionally, branching corals present the most complex and interesting reef features to study from the point of view of image-based, 3D reconstruction. The piece of coral under consideration has been digitised with a laser scanner for accurate assessment and validation of computed reconstructions. By utilising the entire data set it has been shown in previous publications that our 3D reconstruction of the piece of coral is within 1 mm of the ground-truth on over > 90% of the surface area, see Fig. 2(b). The primary focus of this research is twofold: 1) what subsets of the 81 images produce the best 3D reconstruction, and 2) of these best subsets, which ones could be obtained by an underwater vehicle surveying a reef environment. The process utilised for traversing from an image set to a 3D point cloud reconstruction of a scene can be broken down into three distinct phases. Specifically, these are local image feature detection and description, recovering the basic structure from motion (SfM) and dense reconstruction through multi-view stereopsis (MVS). Coincidentally, each relate to distinct software elements in the reconstruction pipeline. Initially, the scale-invariant feature transform (SIFT) algorithm is used to extract common local features