[sparks]

@warren
I think you misunderstood and that we are agreeing. Maybe I wasn't clear but I was trying to say if a read was mapped in first pass to the wrong location or incorrectly aligned (i.e. with mismatches when it should have had an indel) then it will stay that way as it doesn't get to the unmapped file.
Using Bowtie(ungapped) then Novoalign and you'll have this problem.
Using Novoalign/Novoalign with a low threshold on the first pass is OK and will produce the same results as just one run at high threshold.
BWA as first aligner should be fine as it has low false positive rate.

[sparks]

I think Heng Li and krobison have it right. You need to take more than the unmapped reads, you need to take pairs that are not properly aligned and alignments with the number of mismatches or score above some limit. The limit will depend on the first aligner used.