For long reads such as 454, Mosaik performs quite well, particularly if you are interested in insertions and deletions--- although pyrosequencing tends to produce a lot of homopolymer runs in the data which can frustrate indel detection.
FreeBayes, which operates on BAM files as produced by most contemporary short-read aligners, can be used to detect SNPs and short (less than read-length) insertions and deletions. (I am the author.)
If you want to obtain the highest-possible quality dataset, I recommend using multiple approaches to obtain your list of variations, and then intersecting this set between different approaches. This can help you to remove variants which are artifacts of a specific computation methods.
Alternatively, you can take a union of different approaches to find the variants which specific algorithms miss, and then filter the resulting dataset using reported variant quality metrics (as found in the VCF files which the variant detectors output).