Indeed, the only difference is they work on RGB space, and the dataset is a bit toy-ish (no offence), as the networks simply need to separate the objects either by color, or a regular texture pattern.
What proposed in this motion grouping paper, is more like on the idea level, which gives an observation that, although objects in natural videos or images are of very complicated texture, and there is no reason a network can group these pixels together if no supervision is provided.
However, in motion space, pixels moving together form an homogeneous field, and luckily, from psychology, we know that any parts of the objects tend to move together.
CUDA is superior to OpenCL. Also NVIDIA provides CuDNN, a proprietary library with very efficient implementation of deep learning primitives. If you want to train models faster, you have to use them.
CS is the craziest of them all. Those should be the easiest to replicate. "Here is the code, here is a manifest of the environment/container/disk image/etc." You should be able to take that and run it and get the same results.
Or are you saying that the code itself is the problem and that they've done the equivalent of "return True" to get the result they want?
In my other comment I mentioned the CS results I've largely struggled to reproduce is because they include enough detail for you to get the gist of how it works, but not enough to avoid going down some rabbit holes. Also, not all publications include code. Many venues don't require it.
Tagger: Deep Unsupervised Perceptual Grouping
https://arxiv.org/abs/1606.06724