The real issue is this: DESeq2 is meant for testing for differential expression of genes, not isoforms. Quantification of different isoforms for the same genes will always have correlated errors, and unless these correlations are explictely accounted for, any test for differential expression is bound to produce loads of false positives.
So, just counting reads for each isoform independently and running DESeq (or any other testing method) is a big no. (See this old post if it is unclear why: A: DESeq and transcript-wise analysis)
Methods like RSEM (and similarly Sailfish, I suppose) avoid double-counting reads for several isoforms, so the problem is not exactly as described in the post above. Still, the errors (i.e., deviations of RSEM's estimates from the true values) will be correlated between different isoforms of the same gene. RSEM is actually able to estimate and report these correlation matrices, and a method that used this additional information would be able to perform correct inference. However, I am not aware of any tool that is designed to process RSEM's error estimates for purposes of comparing isoform expression.
So, if you are just interested in comparing on the level of genes, use a method that quantifies genes, not isoforms.
If you are interested in comparing on the level of isoforms, use a method that is able to estimate these correlations and account for them in testing. Tools like BitSeq or cuffdiff integrate both the quantification and testing in a single tool or tool-chain, precisely for the reason that the correlation estimates must be passed from the quantification to the testing stage to avoid false positives, and this is typically not possible if you combine a quantification tool from one author with a testing tool from somebody else.
If, however, you are willing to follow our argument that testing on the exon level instead of the isoform level not only avoids all these difficulties but also provides more interpreatble results, use our DEXSeq package (and read our paper on it).