Loop-closure detection is an important component of autonomous navigation and visual localization.In recent years, with the development of deep convolutional neural networks, loop-closure detection based on deep convolution descriptors and manual design descriptors have been applied in various systems. The engineering performance of the mobile robot was quantitatively evaluated and compared by using the image matching based on convolutional neural network and global localization algorithm. From the designed experimental methods, the advantages and disadvantages of the main descriptors in various scene applications were quantified and compared. It is found that the convolution descriptors generated based on AlexNet are the best in each experiment. It provides a quantitative reference selection standard for image matching and its global localization.