Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Person Re-ID in 2025: Supervised, Self-Supervised, and Language-Aligned. What Works?

Lakshman Balasubramanian

TL;DR

This paper benchmarks three ReID training paradigms—supervised, self-supervised, and language-aligned foundation-model–based approaches—across 11 models and 9 datasets to assess cross-domain robustness. It finds that supervised models excel within their training domain but collapse under domain shifts, while language-aligned models (e.g., SigLIP2) offer strong cross-domain generalization even without explicit ReID training, with hybrid CLIP-ReID methods delivering the best cross-domain performance overall. Self-supervised methods underperform in zero-shot ReID, suggesting a lack of identity-level semantic structure. The authors argue for hybrid strategies that combine discriminative supervision with semantic robustness, and they highlight future directions in text-guided retrieval, privacy-preserving training, and model efficiency to make robust ReID deployment practical.

Abstract

Person Re-Identification (ReID) remains a challenging problem in computer vision. This work reviews various training paradigm and evaluates the robustness of state-of-the-art ReID models in cross-domain applications and examines the role of foundation models in improving generalization through richer, more transferable visual representations. We compare three training paradigms, supervised, self-supervised, and language-aligned models. Through the study the aim is to answer the following questions: Can supervised models generalize in cross-domain scenarios? How does foundation models like SigLIP2 perform for the ReID tasks? What are the weaknesses of current supervised and foundational models for ReID? We have conducted the analysis across 11 models and 9 datasets. Our results show a clear split: supervised models dominate their training domain but crumble on cross-domain data. Language-aligned models, however, show surprising robustness cross-domain for ReID tasks, even though they are not explicitly trained to do so. Code and data available at: https://github.com/moiiai-tech/object-reid-benchmark.

Person Re-ID in 2025: Supervised, Self-Supervised, and Language-Aligned. What Works?

TL;DR

This paper benchmarks three ReID training paradigms—supervised, self-supervised, and language-aligned foundation-model–based approaches—across 11 models and 9 datasets to assess cross-domain robustness. It finds that supervised models excel within their training domain but collapse under domain shifts, while language-aligned models (e.g., SigLIP2) offer strong cross-domain generalization even without explicit ReID training, with hybrid CLIP-ReID methods delivering the best cross-domain performance overall. Self-supervised methods underperform in zero-shot ReID, suggesting a lack of identity-level semantic structure. The authors argue for hybrid strategies that combine discriminative supervision with semantic robustness, and they highlight future directions in text-guided retrieval, privacy-preserving training, and model efficiency to make robust ReID deployment practical.

Abstract

Person Re-Identification (ReID) remains a challenging problem in computer vision. This work reviews various training paradigm and evaluates the robustness of state-of-the-art ReID models in cross-domain applications and examines the role of foundation models in improving generalization through richer, more transferable visual representations. We compare three training paradigms, supervised, self-supervised, and language-aligned models. Through the study the aim is to answer the following questions: Can supervised models generalize in cross-domain scenarios? How does foundation models like SigLIP2 perform for the ReID tasks? What are the weaknesses of current supervised and foundational models for ReID? We have conducted the analysis across 11 models and 9 datasets. Our results show a clear split: supervised models dominate their training domain but crumble on cross-domain data. Language-aligned models, however, show surprising robustness cross-domain for ReID tasks, even though they are not explicitly trained to do so. Code and data available at: https://github.com/moiiai-tech/object-reid-benchmark.
Paper Structure (59 sections, 11 equations, 4 figures, 2 tables)

This paper contains 59 sections, 11 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Formal Definition
  4. Evaluation Metrics
  5. Mean Average Precision (mAP)
  6. Cumulative matching Characteristic (CMC)
  7. Overview of ReID Methods
  8. Supervised ReID Training
  9. Domain Generalization Approaches
  10. Fine-Tuning Foundation Vision Encoders
  11. Model Paradigms Under Study
  12. Supervised Learning
  13. Identity Classification Loss
  14. Triplet Loss
  15. Self-Supervised Learning
  16. ...and 44 more sections

Figures (4)

  • Figure 1: Cross-domain performance degradation. Supervised models excel on surveillance but catastrophically fail in-the-wild. Language-aligned models show the opposite pattern. Surveillance = avg(MSMT17, Market, Duke, CUHK03). In-the-Wild = avg(LasT, CelebReID).
  • Figure 2: Cross-domain generalization patterns. CLIP-ReID retention rate calculated as (target_mAP / MSMT17_mAP) × 100. Performance drops sharply for out-of-domain datasets (LasT, Celeb). SigLIP2 shows opposite pattern with higher retention on diverse datasets, with values $\geq$ 100% indicating better performance than baseline.
  • Figure 3: Best model performance by paradigm across dataset types. Supervised models dominate on their training domain (MSMT17) but collapse on in-the-wild datasets. Language-aligned models show the opposite pattern, excelling where supervised models fail.
  • Figure 4: Model size vs. average performance across all datasets. Larger models do not guarantee better performance. Supervised models (blue circles) achieve highest performance despite modest size. PE-Core (671M, green triangle) underperforms despite being the largest model.