The article does not use LSH as it is typically used.
Typically, LSH is for low dimensions - you split the space to cells by the hash value, use the hash value as an index, and look up one to a few values for every key (so you have some overlap).
In contrast, in the article, we use the value not as a hash but to construct a metric, approximating the original metric space. Then, we do a full scan with filtering by the approximate distance. It speeds up the search because a low amount of data is scanned, but it is not like using the hash for direct lookups.
You can name it differently, e.g., "quantization into bit-vectors".