Source of Variant Loss: Assembly or Pileup?

Question

[pld]

I've been comparing some alignments with BWA to the Roche assembler, and I'm finding that with BWA (BWASW) I'm able to capture all of the SNPs found by the Roche assembler, but not the single base deletions ocuring in homopolymers that the Roche assembler finds.

Where would be the best place to focus on tweaking? The BWA parameters, or trying to see if the problem is that samtools mpileup ends up filtering it out?