This note clarifies that the entity referred to in parts of the label-distribution-learning (LDL) literature as S-JAFFE is not distinct from JAFFE, the Japanese Female Facial Expression dataset. It does not introduce substantively new images or new primary annotations beyond those already contained in the original JAFFE resource; rather, it is a task-specific reformulation of JAFFE based on information already present in the original dataset (Geng, 2016; Lyons et al., 1998). We argue that use of the label S-JAFFE has contributed to citation drift and scholarly misattribution by encouraging later authors to treat a derived representation as a distinct dataset (for example, Xu et al., 2019). To reduce further credit displacement, future work should discontinue the label S-JAFFE, refer instead to JAFFE, and cite the original JAFFE dataset according to the dataset terms of use.
Michael J. Lyons (Thu,) studied this question.