Szczegóły publikacji
Opis bibliograficzny
Quality-gated mask attention for exercise recognition under speed shift / Kamil WARCHOŁ, Bogdan KWOLEK // W: Artificial Neural Networks and machine learning - ICANN 2026 : 35th International Conference on Artificial Neural Networks : Padua, Italy, September 14–17, 2026 : proceedings , Pt. 4 / eds. Luca Pasa, [et al.]. — Cham : Springer Nature, cop. 2027. — ( Lecture Notes in Computer Science ; ISSN 0302-9743 ; 17092 ). — ISBN: 978-3-032-38406-5; e-ISBN: 978-3-032-38407-2. — S. 734–746. — Bibliogr., Abstr. — Publikacja dostępna online od: 2026-09-14
Autorzy (2)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 170129 |
|---|---|
| Data dodania do BaDAP | 2026-09-22 |
| DOI | 10.1007/978-3-032-38407-2_58 |
| Rok publikacji | 2027 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Wydawca | Springer |
| Konferencja | International Conference on Artificial Neural Networks 2026 |
| Czasopismo/seria | Lecture Notes in Computer Science |
Abstract
As recently shown, video transformers that condition on human segmentation masks can improve exercise recognition, but mask quality varies frame to frame because automatic segmenters often produce empty, fragmented, or spatially drifted masks, and these failures worsen under execution-speed domain shift. Standard cross-attention treats every mask equally, propagating segmentation noise into the temporal representation. In this work, we introduce RMCAT (Reliability-aware Mask-Conditioned Attention Transformer), which equips a mask cross-attention backbone with per-frame quality awareness. Analytical geometry features quantify mask area, centroid displacement, and boundary fragmentation without trainable parameters. A differentiable reliability gate interpolates between the cross-attended token and a video-only fallback, while an auxiliary head supervises mask quality via a proxy target during training. On a cross-speed exercise benchmark the best variant reaches 88.2 % accuracy and 85.6 % macro-F1, a +2.3%-point F1 improvement over a strong retrained baseline. Consistent gains on the publicly available MM-Fit benchmark with fully automatic YOLO masks confirm that the approach transfers across mask-noise profiles.
