Beyond Deterministic Alignment: Probabilistic VisionLanguage Representations for Pedestrian Re-Identification

Authors

  • Hiroshi Tanaka Graduate School of Engineering, The University of Tokyo, Bunkyo-ku, Tokyo 113-8656, Japan Author
  • Yuki Nakamura Graduate School of Engineering, The University of Tokyo, Bunkyo-ku, Tokyo 113-8656, Japan Author
  • Kenji Sato Graduate School of Engineering, The University of Tokyo, Bunkyo-ku, Tokyo 113-8656, Japan Author
  • Ayumi Kobayashi Graduate School of Engineering, The University of Tokyo, Bunkyo-ku, Tokyo 113-8656, Japan Author

DOI:

https://doi.org/10.71465/mrcis210

Keywords:

Pedestrian re-identification, uncertainty modeling, probabilistic learning, vision language models, autonomous driving

Abstract

Uncertainty estimation plays an important role in perception tasks for autonomous  systems. Motivated by CLIP-based uncertainty modal modeling, this study formulates pedestrian  representations as probabilistic embeddings rather than fixed feature vectors. A variational  learning strategy is employed to model both data-dependent and model-related uncertainty across  visual and textual modalities. The approach is evaluated on three pedestrian re-identification  benchmarks containing over 95,000 identities and 780,000 images. Performance is compared with  deterministic baselines, including ResNet-based ReID, Transformer-based ReID, and CLIP derived embedding models. Results show average gains of 3.1% in rank-1 accuracy and 3.7% in  mAP, along with a reduction of expected calibration error by 18%–24% in occlusion-heavy  scenarios. 

Downloads

Published

2026-01-31

Issue

Vol. 6 No. 1 (2026): MULTIDISCIPLINARY RESEARCH IN COMPUTING INFORMATION SYSTEMS

Section

Articles

License

Copyright (c) 2026 Hiroshi Tanaka , Yuki Nakamura , Kenji Sato , Ayumi Kobayashi (Author)

Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 International License.

All articles published in the Multidisciplinary Research in Computing Information Systems are licensed under an open-access model. Authors retain full copyright and grant the journal the right of first publication. The content can be freely accessed, distributed, and reused for non-commercial purposes, provided proper citation is given to the original work.