The purpose of RSEM, over a simplier quantification methodology, such as counting reads that overlap the genomic region of the gene, is to estimate which isoforms of a gene reads are coming from.Methods like RSEM, Salmon and Kalisto do this using a statistical approach called expectation maximalization (EM). They basically choose the transcript composition that best explains the data.
In the highlighted transcript in your IGV screen short notice that what sets it apart from other transcripts is its 6th exon. While other transcripts do have exons that overlap this exon (and indeed, shared a 5' end), the 3' exon junction varies. Now note that very few reads map to the region of this exon, and many of those are not compatible with having come from that exon (they overlap the intron as well as the exon). When choosing which of the transcripts - the highlighted one, and the three below it, are the most likely source of the few reads that do overlap this region, RSEM is effectively looking at the ratio of 5' spliced reads to 3' spliced reads - it has probably found that on the evidence of the 3' spliced reads, the other transcripts better account for the reads mapping to the region of the exon than the highlighted transcript. (the most obvious combination would be that there are no reads that support the 3' end of the highlighted transcript, and the number of reads that support the 3' end of the other transcripts sum to the number of reads supporting the shared 5' junction).
Its worth noting that this sort of EM estimation of transcript expression is quite impressive, but it is not 100% accurate. Most these estimates come with error bars, and estimate of transcript expression is definitely not as good as at the gene level.