We will later on have a class that will be responsible for feeding data into the model during the training.
We could just feed the raw images, but we won't stop here.
The idea is to have some functions that will modify the images and annotations to increase the accuracy of the model.
Here are some function ideas (higher is highest priority):
[x] image / annotation flip : flipping images and steering (and maybe some other values that can be mirrored)
[x] random noise : add some noise to the image.
[x] random brightness : change the brightness of the image.
[x] random shadows : change the brightness of a region on the image.
[x] color shift : maybe vary a bit the colors on the image.
[x] simulate a car in front : add a randomly placed image of back of cars on the image (maybe also offset the steering to be able to overtake). The size of the car will also vary.
We will later on have a class that will be responsible for feeding data into the model during the training.
We could just feed the raw images, but we won't stop here. The idea is to have some functions that will modify the images and annotations to increase the accuracy of the model.
Here are some function ideas (higher is highest priority):