Bottled Rotation

Brian works at a beer factory where he processes pictures of the bottles before they are shipped out.

He noticed that the pictures were often taken with different degrees of rotation, and the Convolutional Neural Network he built wasn't ready to handle this.

Which of the following approaches could Brian propose to handle rotation in the pictures?

Adding a data preprocessing step to properly rotate every image before giving the data to the model.

Clustering rotated images before sending them through the model and using this as a separate class.

Include rotated versions of the images in the training data to build some rotation invariability into the model.

Correctly configure the network since Convolutional Neural Networks are inherently rotation invariant.