What you see is a three-dimensional visualization of tracks clustering (use your mouse or touch to rotate the picture!).
To be more precise, there is a number of track segments that were left after preprocessing the data with ML classification techniques. Out goal here is to complete the job and reconstuct tracks that were left by the particles.
Good clustering meets all the above goals.
While there is no comparison of clustering techniques used (DBSCAN algorithm was used in all the cases), I'm focusing on comparig different ways to introduce a distance between elements. It's extremely important to choose metric wisely, since it is the only information available to clustering alogrithms.
As you may learn from this example, good choice of distance (metric) is crucial and may be a choice between a complete win and an absolute nonsense. Chosen metric should incorporate your prior knowledge about the application you're working with.
You can read more details on how do we collect such information, and why are such studies are important in the following posts: