Information on the spatial location of road facilities such as traffic signs is one of the basic element of urban three-dimensional (3D) modeling, and it is also an essential part of road facility maintenance and management. To this end, an automatic extraction scheme for small traffic signs based on mobile measurement data is proposed herein. Based on the improved convolutional neural network SlimNet model, small cross-marks on panoramic images are detected, and a 3D target geolocation based on depth maps is proposed. A distance assessment method based on the center point is used to extract the diagonal of the target. Measured data of the three types of small traffic signs are selected to verify and analyze the proposed method. Experiment results show that the average accuracy of the SlimNet model is 4.2 percentage higher than that of the classic VGG16 (Visual Geometry Group 16) model. Using the proposed geographic positioning and vectorization scheme, the recall rate and accuracy rate of the three types of targets in the experimental area reached over 86%, proving the effective feasibility of the overall scheme. Furthermore, the proposed method provides a new idea for an accurate 3D spatial geolocalization of urban multi-class targets.