his dataset was derived from the original Stanford Cars Dataset, which can be downloaded here: http://ai.stanford.edu/~jkrause/cars/car_dataset.html

Since the year level was not accurately labeled when the Stanford Cars Dataset was developed (which can be found in Stanford's research publications on this data), the year level of the SC was dropped. The duplicate class labels due to the removal of the year level were simply merged together.

All images in this dataset are pre-processed, meaning all bounding boxes information was used in cropping the images so that the object of interest is fully focused within the image. The training and the testing set are split with 50:50 ratio, which was how Stanford did it when this data was published. Further pre-processing is needed if this train/test split ratio has to be altered.

The total class count is 189 in this dataset, which is only 7 less than the original class count of 196.