Original title: Orthogonal Ensembles and Tested Explanations for Performer-Independent Body-Motion Emotion Recognition
Naoto Nishida, Yoshio Ishiguro · September 2026
Recognizing emotions purely from body motion is very difficult, especially when testing on people the computer has never seen before. Guessing by chance yields only 8.3% accuracy, and a standard baseline model scored just 25.73 +/- 4.03%. To fix this, researchers grouped eleven different models that make different kinds of mistakes, averaging their predictions. This combined group raised performance to 36.80 +/- 4.00%. Next, the researchers tested why the models worked by masking body parts and making edits. They proved the models make decisions by watching specific body regions. These regions matched established movement theories, known as Laban Movement Analysis, much better than traditional physics-based movement measurements. However, the tests revealed a downside: the models spread their focus diffusely across entire time windows instead of pinpointing specific moments.
Imagine you're trying to guess what mood someone is in just by watching them walk, without hearing their voice or seeing their face. Normally, you might focus on facial expressions, but here you only have their body movement. In this paper, the researchers use computer models that focus on specific parts of the body, like arms or legs, and these models together get better at guessing emotions by combining their different focuses, much like how different people might notice different details when observing movement.
This improves AI systems that interpret emotions from body language, potentially aiding human-computer interaction or therapy. By clarifying which body parts matter, it makes the technology more transparent and trustworthy.
The work is limited to acted emotions from skeleton data, not spontaneous emotions in real life. It does not address the timing of emotions within a movement sequence. The overall accuracy is still relatively low in a challenging evaluation setting.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Orthogonal Ensembles and Tested Explanations for Performer-Independent Body-Motion Emotion Recognition — Naoto Nishida, Yoshio Ishiguro
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.