In this study, the effect of the skip connections, which are seen in fully convolutional networks, on object localization is analyzed. For this purpose, a local data set for plate detection is created. Experiments are carried out using this data set. Due to the small size of the image set, data augmentation method is used to overcome the danger of over-fitting. The learning rates of the first layers are frozen for analysis and fine-tuning is applied to only the last layer and deconvolution layers. The results obtained are compared with the results of other image sets. The results indicate the importance of the information provided by the skip connections on object localization.