I agree that the Matsukura & Hollingworth paper is a very interesting read, but it also has severe limitations that I will briefly sketch below. One of the biggest problems with visual sensory memory tasks is that people initially are terrible at using retrospective cues. In fact, in the first half hour of testing people generally show little to no evidence of partial-report superiority over whole-report, working memory conditions. This does not mean that people represent a limited amount of information in visual sensory memory initially, but just that transfer of sensory memory information to a reportable stage requires practice. The classic literature on iconic memory always tested people for hours if not days on the same task, while nowadays the entire experiment is usually finished within half an hour. Consider that Matsukura et al. used 8 practice trials and 340 normal trials. This would easily fit into a 60-minute test session and thus would lead to a gross underestimation of partial-report superiority effects. In fact, when Matsukura et al. do train subjects longer (experiment 4), their results become quite similar to the results reported here.

Moreover, the Matsukura paper misrepresents two aspects of the experimental design used in this paper:
1) They state that the co-linear alignment of the stimuli in experiment 1 boosts performance in a substantial way and that when four orientations are used (in experiment 3) capacity suddenly declines from 16 to 5.5 items. However, in experiment 1, capacity is 16 with a set size of 32 objects while capacity was assessed with a set size of only 8 objects in experiment 3. Therefore, this is not a reasonable comparison (max performance of 32 in exp1 vs 8 in exp2).
2) Related to the previous point, all objects except for one changed in experiment 1, so if people would use this general grouping principle, they would always respond yes on each trial (leading to bad performance). This procedure was done to prevent grouping strategies. In general, grouping only really helped performance on working memory trials (see experiment 3), not on retro-cue trials.