Object detection for container corner detection

Abstract

In this bachelor’s thesis different deep neural networks for object detection are studied. The networks are used to detect corners of containers in harbor. The studied network types are Single Shot Detector, Faster Region based CNN and RetinaNet with MobileNet and ResNet as base networks. The dataset used is built as part of the study and the networks are implemented using the Tensorflow Object Detection API. The detection performance of the models are compared using average precision without K-fold validation on a single test set. From the results we can see that the that using a pretrained network reasonable detection performance can be achieved using SSD or Faster R-CNN. RetinaNet however performs noticeably worse than the other methods.