I think that 3d reconstruction and related computer vision is going to take over vision applications that have been so far handled by 2d methods. There will be

I think that 3d reconstruction and related computer vision is going to take over vision applications that have been so far handled by 2d methods. There will be more emphasis on well-factored understanding. That 3d understanding will be leveraged by question answering systems with a type of visualization of scenarios. Abstractions will be built on this grounded foundation. Imagine full 3d understanding of a video and association with captions. Now say you train it on 1 million videos. Combining well-factored and fully understood spatial/visual information with text for training along with recent techniques like transformers will finally demonstrate true natural language comprehension for a variety of domains. Look at Hinton's recent paper on using transformers with capsules etc.