Dataset notes

Challenge

The dataset is relatively challenging, with a small amount of data per-class. Consider the task of predicting the following: a Golf 7 & a Golf 7.5 An issue with car images is that a car model (say Golf 7) can have many exterior visual options, for instance, different headlight designs. In the above example, the golf 7.5 has so-called "Xenon" headlights whereas the 7 has standard "Halogen" headlights. Both options are available on either the Golf 7 or 7.5 and can visually differentiate the image substantially. Across the dataset, an attempt has been made to keep the visual "options" consistent across each class by including images of the base car model only, that has no exterior options.

Car pose

When initially starting this project, front, rear and side profile images were collected. Side profile images were disregarded after some initial testing, as the model did not generalize well. These data points were deemed to not be discriminative and once removed, the model generalized better. These data points will be re-added once the model can generalize well on frontal and rearward images. Another option is to add the task to the classifier: y = [pose, model] where pose could be [front, rear, side-profile].

RGB woes

An investigation that needs to be made is what the count of car colors are per class - the model could be over-fitting to this feature. An interim fix could be to convert the images to greyscale and broadcast the greyscale channel to match an input depth of 3. Once this has been completed and depending on the results, multi-task learning (y = [car model, color]) could be promising, to update the model's learning outcomes.

Class balance

Classes are relatively balanced with a count of 60, 68, 83, 89 for Golf 5, 7.5, 7 and 6 respectively. Golf 7 and 7.5 could be concatenated into one class.

Car location

We can see in the above image of the Golf 7.5, another car is present in the background. This could make the classification task convoluted, and a detection problem might be more suitable for the given dataset.