Close the cross-domain valence/arousal transfer gap

Close the gap between the native cross-domain valence/arousal head and post-hoc categorical-channel or frozen-feature approaches for facial affect models, despite the absence of comparable dimensional valence/arousal labels for children.

Background

The paper finds that categorical facial-expression labels are substantially more recoverable under domain shift than continuous valence/arousal estimates. For EmotiEffLib, the native jointly optimized valence head achieves a valence concordance correlation coefficient of 0.375 on the adult ABAW hold-out, whereas regression from frozen features reaches 0.218, mapping categorical posterior probabilities to valence/arousal reaches 0.250, and backbone fine-tuning reaches 0.276.

The authors report that scaling the categorical-to-dimensional mapping to 138,000 adult frames plateaus at 0.335 and that adding degraded predicted action units provides no improvement. Because available child datasets lack dense, per-frame valence/arousal annotations comparable to adult benchmarks, the paper leaves unresolved how to close the remaining cross-domain gap and transfer the categorical-channel advantage to dimensional affect estimation in children.

References

The dissociation holds, yet closing the gap to the native cross-domain V/A head is the open problem: more adult data does not close it, and the binding constraint is the absence of comparable dimensional child labels.

— Whose Face Is It Anyway? A Multi-Model Audit of Facial Affect Recognition on Children, and Why the Gap Is the Head, Not the Features  (2610.08279 - Hallmen et al., 6 Oct 2026) in Section 6, Method: Recalibrate the Head, paragraph “Valence/arousal is harder, and reveals a dissociation”