TMM normalization on background regions aims to correct for composition biases. Specifically, when you get increased binding in one library, more reads are spent in sequencing the increased enrichment of fragments. This means that you have fewer reads to go around for the rest of the genome. Spurious differences may then be observed when this library are compared to other libraries.
The idea with the normalization is to count reads across background bins, and to equalize the coverage across the background between libraries. This assumes that background coverage should be the same between libraries; any systematic differences must be caused by composition bias. Note that composition biases can occur in both TF or histone mark experiments, so the decision to use it isn't really governed by the type of experiment.
The real choice that you should be concerned about is whether you should TMM normalize on the (high-abundance) windows directly. This assumes that most windows are not DB, such that any systematic differences in window coverage between libraries are removed. The idea is to eliminate spurious differences caused by variable IP efficiency between libraries. However, this will also eliminate any large-scale DB between libraries, e.g., if binding increases across many sites in one condition.
The choice between this window- and background-based methods depends on whether you expect to see overall changes in binding intensity between your libraries. If so, you should use the background-based method, as this will preserve the systematic differences for later detection. If not, any differences are assumed to be technical, so the window-based approach should be used. An intelligent choice usually requires knowing some biological context for your study.
