The dataset consisting of 6701 PNG images (64x64 pixels) of straight arrow type road markings with different dimensions and orientations, together with the angle it forms with respect to the vertical axis (azimuth), has been constructed in the framework of the SROADEX project to train a regression neural network to calculate the azimuth of other straight arrows that can be identified in high resolution aerial orthoimages. The dataset was created by labeling with the labelme tool the straight arrows on orthoimage tesserae of 256x256 pixels. After semantic labeling by drawing an arrow on the orthoimage, the data was automatically processed to calculate the azimuthal angle. The procedure followed is as follows: 1.- The vertices of the generated arrow-shaped polygon have been extracted. 2.- Clusters of nearby points have been generated, with a minimum of 2 points and the cluster with less points and the one with more points has been identified. 3.- For the clusters, the centroid has been generated, preserving the information of the number of vertices that define the cluster. 4.- Then a vector has been generated with origin in the centroid of the arrow with less vertices grouped and as end the centroid with the highest number of vertices. 5.- Finally, the azimuth of this vector with respect to the ordinate axis has been calculated, The images have been automatically cropped with a constant size of 64x64 pixels taking an extension greater than the occupied of the arrow in the scene. The shared dataset consists of a compressed file with the images and a text file in CSV format with the names of the images and the normaliced [0..1] Azimutal angles of the straight arrows contained in the images.

Grant PID2020-116448GB-I00, funded by the Agencia Estatal de Investigación, Ministerio de Ciencia e Innovación (ES)