A video dataset for hadal snailfish along with the benchmark

Abstract

Application of deep learning technology for deep-sea ecological studies is still in its infancy stage especially in the field of automatic taxonomic identification and statistics. In this study, we created a novel dataset containing annotated videos for the rare species of hadal snailfish inhabiting in depth below 6000 meters, and conducted control experiments by combining models of different specifications and adding different attention mechanisms. We successfully generated a set of benchmark test data from a quantitative perspective. In addition, based on out of set data with completely different data distributions from the training and validation sets, the generalization ability of the model trained on the new dataset in real-world scenarios was qualitatively analyzed. Other researchers can continue to expand and supplement the dataset based on our benchmarks, or directly apply our results to actual deep-sea videos collected, and accurately identify and capture deep-sea snailfish in the videos. With this deep learning video processing technology, distribution pattern and biodiversity of the deep-sea organisms will be accomplished efficiently.