In the MOT-20 dataset’s ground truth, a visibility score is given which denotes how much of the object is visible. I want to crop objects from this dataset and input it to ResNet-34 with a visibility score. The goal is to learn to predict the visibility score when an object is given. To perform this task, what features could be added to ResNet so that it can learn if an object is occluded and its occlusion score? How should I design the loss function?