Training data with static occlusions

I am about to start training with DeepLabCut on a dataset of cattle passing through a race, the goal is to identify key points on the cattle which will be used in conjunction with depth cameras to calculate biometric measurments.

I have a question regarding the training process, because my dataset involves cattle passing through a race, the animal is partially occluded by this static race. See this image as an example:

Because this is a static environment, i.e the cameras, race will never move, i had an idea to use a image mask to remove some of the race and external environment from the image leaving just the animal inside the race, see the following as an example:

Would this typically yield better results using DeepLabCut, or should i simply train on the full images?

Any advice is much appreciated.

I would train on the full image. If you want the network to “guess” where occluded parts are, just label them anyhow, and it will learn to predict the “nose” even when blocked by the rail in the cattle shoot.