icassp2027-experiment/results/easycom_omni_floor_screen_2026-09-16/README.md

EasyCom Omni floor-event screening

Protocol erratum. Dense and raw-boundary metrics remain valid. All opportunity-conditioned ENTRY/INT values below used the superseded v1 loader, which could cross an intervening ENTRY competitor and selected the latest onset in ambiguous multi-competitor INT cases. Retain them for provenance only; do not use them for model selection or manuscript claims before a v2 rerun.

This record freezes the first comparison of EgoSpeak, Omni G+A, Omni G+A+pooled-V, and the three-seed current-frame F1 Frame-set Attention screen on EasyCom Sessions 1--3. Values were produced on the remote experiment host and supplied by the user on 2026-09-16. Generated predictions and per-event CSVs remain on that host and are not committed here.

Evaluated layers

1. Dense future activity uses causal-valid framewise AP/mAP over ten horizons.
2. Raw wearer endpoint and aggregate-other-boundary AP are threshold-free timing diagnostics; they are not semantic EOT/INT labels.
3. Opportunity-conditioned ENTRY and INT/YIELD AP is a historical v1 screen that requires rerunning with the corrected v2 competing-risk protocol.

All opportunity runs use identical denominators: 163 ENTRY positives and 335 hard negatives, with 81 invalid/post-outcome cases; and 28 INT positives and 109 hard negatives, with 5 invalid/post-outcome cases. Positive prevalence is 32.731% for ENTRY and 20.438% for INT.

Main results

Model	0.2-s mAP	Mean mAP	Boundary response AP	Boundary yield AP	ENTRY AP	INT/YIELD AP
EgoSpeak	74.204	61.193	29.989	41.078	52.487	20.135
Omni G+A, seed 0	81.723	65.109	28.840	47.873	43.886	21.535
Omni G+A+pooled-V, seed 0	82.027	65.400	29.601	48.344	46.022	24.266
F1, three-seed mean	81.777	64.951	28.423	48.641	41.561	22.991

F1 standard deviations are 0.156 points for 0.2-s mAP, 0.095 for mean mAP, 0.480 for boundary response AP, 0.391 for boundary yield AP, 0.581 for ENTRY AP, and 0.202 for INT/YIELD AP.

The raw endpoint diagnostic, available for the three seed-0 baselines, changes from EgoSpeak to G+A to G+A+pooled-V as follows: onset AP 24.491/25.960/26.019 and offset AP 38.024/45.251/45.371. Thus most generic offset/yield-timing gain comes from G+A; pooled visual input adds almost nothing to raw transition timing.

Historical v1 interpretation

• G+A strongly improves dense forecasting and generic wearer-offset/yield timing over EgoSpeak, but not reviewed semantic opportunities. In particular, its INT AP remains close to prevalence and its ENTRY AP is lower than EgoSpeak's.
• Adding pooled-V over G+A gives only +0.304/+0.291 points in 0.2-s/mean mAP, +0.059/+0.120 in raw onset/offset AP, and +0.761/+0.471 in boundary response/yield AP. The larger but still preliminary changes are +2.135 ENTRY AP and +2.732 INT AP.
• F1 is stable across seeds but does not improve on the historical pooled-V run: its mean changes are -0.250/-0.449 points in 0.2-s/mean mAP, -1.178 in boundary response AP, +0.297 in boundary yield AP, -4.461 in ENTRY AP, and -1.275 in INT AP. The hoped-for pattern of unchanged dense/timing performance with improved semantic opportunity ranking is absent.
• Current F1 attends to a current anonymous visual person set using G+A. It has no per-person temporal history, person-conditioned audio, or verified identity continuity. This screen therefore does not support promoting F1 as the paper's main floor-reasoning method.

Hard limitations

• The opportunity labels are provisional two-pass LLM-assisted annotations, not human-reviewed confirmatory labels.
• The old G+A and pooled-V runs are seed-0 fixed-four-epoch models trained on Sessions 4--12 without development selection. F1 trains on Sessions 4--11 and uses Session 12 for checkpoint selection. Their comparison is historical, not a protocol-matched attribution.
• F1 needs a three-seed pooled-V control trained and selected with exactly the same train/dev protocol before a confirmatory conclusion.
• All matched-recall trigger/FPR values produced in these exploratory commands selected thresholds on Sessions 1--3 and are excluded from this record.
• Aggregate others activity cannot establish speaker identity, addressee, or semantic interruption intent.

Exact aggregate values are stored in summary.json and metrics.csv. The earlier EgoSpeak-only record is easycom_egospeak_three_level_eval_2026-09-16.