Towards observation condition agnostic fauna detection and segmentation in seafloor imagery for biomass estimation

The performance of automated object detection and segmentation in marine imaging applications is sensitive to hardware and environmental factors that result in a large variability in the appearance of subjects in images. This paper investigates physics based scale normalisation, lens distortion normalisation, and data augmentation techniques to overcome this, working towards a condition agnostic object detection system. A total of over 700 rockfish in images taken from different altitudes using different camera equipped Autonomous Underwater Vehicles at the Southern Hydrates Ridge (depth 780m) are used to train and test object detection and segmentation using Mask R-CNN. Images taken from low altitudes of 2m achieve a maximum mean average precision (mAP) score of 97.42%, and images taken from high altitudes of ~6m achieve a maximum score of 87.4% when object detection and segmentation is trained and tested on images taken from the same altitudes. When transferring knowledge across different imaging conditions, a mAP score of 87.7% is achieved when transferring knowledge from high to low altitude datasets, and 49.6% when transferring from low to high altitudes. In both cases, significant gains in performance is seen when the images used are scale normalised. The results indicate that increasing the pixel resolution, or the size an object appears within the image, benefits learning regardless of the optical resolution images are taken at, and this should be carefully considered in future object detection and segmentation studies. We also describe a novel method to estimate biomass distribution from the segments output by modern machine learning algorithms that can be easily adapted for different morphospecies