Pakistani Traffic-sign Recognition Dataset

The data collection was carried out over several months and across several cities including but not limited to Quetta, Islamabad and Karachi, Pakistan. Ultimately, the number of images collected as part of the Pakistani dataset were, albeit in a very small quantity. The images taken were also distributed across the classes unevenly, just like the German dataset. All the 359 images were then manually cropped to filter out the unwanted image background data. All the images were sorted into folders with names corresponding to the label of the images. A python script was used to rename the images from 0-359 in alphabetic order of their class labels (so that images from ‘Bridge Ahead’ were named 0-12 whereas images from ‘Zigzag Road Ahead’ were named 346-359). The sizes of these cropped images were non-uniform but, a CNN only accepts input images with uniform dimensions, whatever they might be. Hence, all the images were then resized to the shape of 32 x 32.