Partial occlusion can be solved through convolutional features.
Complete occlusion would need to store more than just t-1 in memory and also use t-2 … That’s the first solution I can think of