The Worldwide Journal of Robotics Analysis, Forward of Print.
Open-source benchmark datasets have been a crucial element for advancing machine studying for robotic notion in terrestrial functions. Benchmark datasets allow the widespread growth of state-of-the-art machine studying strategies, which require giant datasets for coaching, validation, and thorough comparability to competing approaches. Underwater environments impose a number of operational challenges that hinder efforts to gather giant benchmark datasets for marine robotic notion. Moreover, a low abundance of targets of curiosity relative to the scale of the search house results in elevated time and value required to gather helpful datasets for a particular activity. Because of this, there’s restricted availability of labeled benchmark datasets for underwater functions. We current the AI4Shipwrecks dataset, which consists of 28 distinct shipwrecks totaling 286 high-resolution labeled facet scan sonar photographs to advance the state-of-the-art in autonomous sonar picture understanding. We leverage the distinctive abundance of targets in Thunder Bay Nationwide Marine Sanctuary in Lake Huron, MI, to gather and compile a sonar imagery benchmark dataset by way of surveys with an autonomous underwater automobile (AUV). We consulted with professional marine archaeologists for the labeling of robotically gathered information. We then leverage this dataset to carry out benchmark experiments for comparability of state-of-the-art supervised segmentation strategies, and we current insights on alternatives and open challenges for the sector. The dataset and benchmarking instruments might be launched as an open-source benchmark dataset to spur innovation in machine studying for Nice Lakes and ocean exploration. The dataset and accompanying software program can be found at https://umfieldrobotics.github.io/ai4shipwrecks/.
发布者：Advaith V. Sethuraman，转转请注明出处：https://robotalks.cn/machine-learning-for-shipwreck-segmentation-from-side-scan-sonar-imagery-dataset-and-benchmark-2/
