Fisheye cameras have recently became very popular in computer vision applications due to their wide field of view. In addition to a better overview of the surrounding area, they enable to capture objects at extremely ...
详细信息
ISBN:
(纸本)9783030243050
Fisheye cameras have recently became very popular in computer vision applications due to their wide field of view. In addition to a better overview of the surrounding area, they enable to capture objects at extremely close ranges. These advantages come at a cost of strong image distortion, which cannot be removed completely maintaining image continuity. This complicates the use of traditional computer vision algorithms, which expect a single image as an input. This paper presents a performance evaluation of neural network algorithms for object detection and segmentation on fisheye camera images. Three approaches are evaluated: semantic imagesegmentation with Fully Convolutional Network (FCN) [13], a fully convolutional approach to instancesegmentation with U-Net [18] and a region-based approach to instancesegmentation with Mask R-CNN [10]. All of these networks successfully solved the task. However, as they were designed to different purposes, each of them has its own strengths and shortcomings. These three approaches are used to perform euro container imagesegmentation task. An image dataset was created in order to train and evaluate these algorithms. Huge part of this dataset was generated artificially, which simplified the task of ground truth labeling. The power of neural networks enable for fast and reliable imagesegmentation. As to our knowledge, this is the first neural networks application for euro container fisheye image detection and segmentation.
暂无评论