Cross-modality person re-identification in video with Mamba-inspired linear attention
Science Progress · 2026-10 · 109 (4)
DOI: 10.1177/00368504261511166 ↗Kurz gefasst
Fachartikel in der Zeitschrift Science Progress, online erschienen am 20.11.2026. Verfasst von Rangwan Kasantikul und 2 weiteren Autor:innen. Beteiligte Einrichtung: Faculty of Information and Communication Technology, Mahidol University, Salaya, Thailand. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0). Gefördert von: National Research Council of Thailand, Faculty of Information and Communication Technology, Mahidol University.
Abstract (Originalsprache)
Purpose
Cross-modality re-identification (Re-ID) such as infrared-to-visible Re-ID requires effective handling of sample variances between two different modalities. Two strategies are modality-aware training of the model, and information extraction from sequence of frames. For this reason, we developed a technique Augmented Adversarial Learning with Temporal MILA (AATM).
Methods
We adopted an enhanced adversarial learning, and the MILA attention model. Greyscale modality derived from coloured video footage is used as an additional modality label for adversarial learning, in addition to using it for training the feature encoder itself. The HITSZ-VCM dataset was used for the evaluation.
Results
The addition of the greyscale input third modality during the training resulted in the improvements of infared-to-visible Re-ID by 3.06% in mean average precision (mAP) and 3.50% in top-1 accuracy compared to using only visible and infrared inputs during training. Under our test conditions, the predictive performance of MILA-based model is similar if slightly better than traditional transformer implementation. However, the MILA-based model lags behind transformer-based model in computational time.
Conclusion
The inclusion of the third greyscale modality for adversarial training contributed to the superior result of the AATM technique when tested on the publicly available HITSZ-VCM dataset. However, the more time-efficient implementations of MILA must be further pursued to justify replacing transformer-based attention in this problem area.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Faculty of Information and Communication Technology, Mahidol University, Salaya, Thailand.
- School of Electrical, Mechanical and Biomedical Engineering, University of Technology Sydney, Ultimo, NSW, Australia.
ORCID-Profile: Worapan Kusakunniran
Förderung
- National Research Council of Thailand
- Faculty of Information and Communication Technology, Mahidol University
Zitieren
APA
Kasantikul, R., Kusakunniran, W., & Wu, Q. (2026). Cross-modality person re-identification in video with Mamba-inspired linear attention. Science Progress, 109(4). https://doi.org/10.1177/00368504261511166
Vancouver
Kasantikul R, Kusakunniran W, Wu Q. Cross-modality person re-identification in video with Mamba-inspired linear attention. Science Progress. 2026;109(4). doi:10.1177/00368504261511166
BibTeX
@article{Kasantikul2026,
author = {Kasantikul, Rangwan and Kusakunniran, Worapan and Wu, Qiang},
title = {Cross-modality person re-identification in video with Mamba-inspired linear attention},
journal = {Science Progress},
year = {2026},
volume = {109},
number = {4},
doi = {10.1177/00368504261511166}
}
EUVIMED