Towards population counting of marine mammals based on drone images

Abstract

In marine science, there is a need for tools for population counting of species. Through this thesis we aim to achieve the follow three objectives: first, briefly discuss the state-of-the-art object detectors that can be used for the detection of porpoises in drone images/videos. Second, test and compare a few stateof-the-art object detectors in both quantitative and qualitative manner. Third, based on our results propose a set of suggestions that can be used for future studies associated with population counting. To answer the second question we compared three state-of-the-art object detection models, two single-stage detectors, and one two-shot detector. The models chosen were the Faster RCNN, YOLOv4 and EfficientDet models, and they were trained and tested on a custom data-set consisting of 7300 labeled images of porpoises, where as 2300 of these were included in the test data set. Through our experiments, we have discovered that YOLOv4 outperforms Faster R-CNN and EfficientDet D1 with detection, where YOLO achieves a recall of 97%, compared to 80% recall with EfficientDet D1 and 75% recall with Faster R-CNN. We also find the average precision 𝐴𝑃@50 values of YOLOv4 to be 0.778, which is greater than EfficientDet D1 with 0.695 and Faster R-CNN with 0.686. Through both qualitative and quantitative methods we discover that both EfficientDet D1 and Faster R-CNN suffers from poor recall especially when porpoises overlap in the images. In the case of Faster R-CNN it misses nearly all detections when the porpoises overlap, but rarely non-overlapping detections. EfficientDet misses a significant portion of the overlapping detections, but also misses a few of the singular. Through examination of the COCO detection metrics, which favor bounding box accuracy, we also show that Faster R-CNN has more precise bounding boxes than YOLOv4 and EfficientDet D1 by comparing the less strict AP@50 values, with the stricter AP@75 and 𝐴𝑃 [.50 : .05 : .95] values. These results imply that a one-stage detection model in YOLOv4 could be used for object detection of porpoises from drone images. Based on the results, a few important areas for further investigation is outlined in the discussion, and a framework was developed which allows marine researches to easily perform porpoise detection from images and videos.