Entrants name the emotion in a motion-capture recording of a single performer. The system sees 24 joints on the body plus one point for where the body is in the room: no face, no sound, no scene.

Each recording is one performer acting one emotion; every model here reads a fixed 64-frame window of it, about half a second. The recordings come from DIEM-A, the Diverse Intercultural E-Motion Database of Asian Performers.

Two people performing the same emotion differ about as much as one person performing two different emotions.

Every score here comes from 10-fold leave-performers-out cross-validation over the 74 training performers, reported as the per-fold mean ± SD.

Every upgrade to a single model failed or made the score worse.

• A supervised-contrastive head cost −11.53 pp.
• Aligning recordings to scene descriptions cost −1.42 pp: the projection collapsed onto performer identity, not emotion.
• Pseudo-labels from a vision-language model failed: on faceless stick figures it ranked the true emotion last of twelve.
• Post-hoc calibration moved the score by 0 to −0.14 pp.

All on the same ten-way performer split. The paper lists eight of these; together they are the ceiling on one better network.

Averaging cancels a mistake only when the members do not make it together, so the members were chosen for how differently they fail.

One larger network was the alternative. The capacity went to these four assumptions instead.

Seven members trained here plus four frozen branches with a linear head each: 12.98 M trainable parameters.

Add the eleven logit vectors and divide by eleven.

The two smaller figures are the same comparison with the members rescaled and then with their magnitudes thrown away, so the gain is not an artifact of one member's logits being larger than another's.

Baseline, best single model, seven members, then eleven: 25.7 → 30.1 → 33.9 → 36.8%.

Each bar is the mean over ten leave-performers-out folds, each dot is one fold, and the pooled out-of-fold score sits beside each bar.

Correlations run over per-sample error indicators on 7,992 out-of-fold clips. The removal costs are pooled out-of-fold Macro-F1 against a base of 36.94%.

Jealousy reads as contempt and guilt as sadness, and top confidence sits near 62%.

The country difference is not a cultural finding.

Japanese and Taiwanese performers score 35.3 and 38.9%. The two groups also differ in when they were recorded and in how each person moves.

Dropping the seven earliest Japanese performers from training moves the score by −0.44 pp.

Row-normalised out-of-fold confusion over the twelve emotions, with the five strongest off-diagonal cells boxed.

The recording-date check gives 36.50 against 36.94% pooled out-of-fold. The same motion also decodes which country a performer is from about 69.8% of the time, which is another reason not to read the country gap as culture.

To claim the model uses a part of the body, we mask it, perturb it, or edit how it moves, and watch whether the prediction moves.

Each check runs on one strong member of the submitted ensemble; the two that carry the most weight are re-run on all eleven.

Masking the parts it ranks high costs more accuracy than masking the parts it ranks low, and the ranking barely moves under noise.

Without these two, the alignment on the next slide would be a picture agreeing with itself.

Masking works over 6 body parts here and the next slide's alignment over 4 regions, two groupings for two checks. At the finer 25-joint resolution the gap grows to +0.199. Both values are 10-fold means.

Laban Movement Analysis is a movement analyst's vocabulary for how a body moves. People wrote its rules; nothing learned them from our recordings.

One strong member. The 12 emotions are not 12 independent observations, so the 95% CI [+0.150, +0.733] is an emotion-block bootstrap and the null permutes regions within an emotion, p < 0.001. Positive for 10 of the 12 emotions (sign test p = 0.039, median ρ = +0.70).

Inside the half-second the model reads, every frame counts about the same, and across the body the head and neck dominate. Order still matters: reverse the clip and every model that reads it in order loses 3 to 5 points.

The window is 64 frames, about half a second at 120 fps. Across frames the important-versus-reverse gap is about +0.002, against +0.124 ± 0.031 across body parts, and frame saliency entropy is 98.7% of its maximum per clip, 99.5% as a class mean. The null is about locating a frame, not about time.