I am currently working on a UNET model that uses Dice loss in order to perform image segmentation. The purpose for the model is to automate a rover by having it better perform object detection. In addition, segmentation would make it easier to calculate the size of an object based on the perspective of the object and the rover’s speed.

The UNET architecture is one of the larger models meaning it takes longer to process images than a simple classifier or GAN. The trade off is that the large amount of feature maps produced inside of it allow the model to train significantly quicker and provide more accurate results. In order for the rover to perform the task of detecting and sizing an object before picking it up, the output of the model must be very precise.

To the right is an image of the version I built.