Udacity has made available an annotated car (and a few other objects) data-set. I’ve used the ‘Dataset1’, annotated by CrowdAI for this project. Please take a look at the ‘Dataset Exploration.ipynb’ jupyter notebook where I’ve explored the same.

In summary,

• the dataset identifies 3 classes: Car, Truck and Pedestrian and also lists bounding box coordinates for each of the objects in datapoint (image), in a CSV file.
• The data-set is uneven across different classes
• It can additionally be noted that a certain view (rear) of cars dominates the rest (side and front)
• The lighting condition is constant throughout the capture and hence we will need data augmentation to help the network generalize better.

If you’d like to play with the training bit, download the data-set and extract it to ‘udacity-object-detection-crowdai/’ in the root of the project folder.