Genuine Open-set re-ID Metric (GOM) Overview

• Genuine Open-set re-ID Metric (GOM) is not formally defined, contrasting with established metrics like mAP and TAR/FAR in open-set recognition.
• Open-set re-ID evaluation combines genuine matching with impostor rejection, highlighting a gap in standardized, principled metrics.
• Recent latent domain expansion and domain discovery methods demonstrate the need for a unified metric to balance domain adaptation with robust performance.

A Genuine Open-set re-ID Metric (GOM) is not defined, mentioned, or implied in the factual record of the provided literature, including (Nitzan et al., 2023) or related domain expansion and latent domain discovery works. Below is a comprehensive, encyclopedia-style synthesis of the relevant context surrounding open-set recognition, re-identification (re-ID), and the current landscape of latent domain expansion and evaluation—but without introducing or hypothesizing the existence, definition, or properties of a GOM beyond attested information.

1. Background: Re-identification and Open-set Recognition

Person re-identification (re-ID) is the task of matching individuals across non-overlapping camera views without direct identity supervision. In classical re-ID, the gallery contains all possible probe identities (closed-set), enabling standard classification or retrieval metrics. In contrast, open-set recognition addresses situations where a query entity may not be present in the gallery, requiring models not only to match known identities but also to reject previously unseen ones.

Open-set re-ID thus conjoins the challenges of traditional re-ID with those of open-set recognition: the need to robustly distinguish between known and unknown identities, handle domain shift, and maintain high retrieval performance across both seen and unseen categories.

2. Metric Evaluation in Re-ID and Open-set Contexts

Standard re-ID metrics include Cumulative Matching Characteristic (CMC) curves and mean Average Precision (mAP), which assume a closed-set (all queries are in the gallery). For open-set scenarios, the conventional metrics become insufficient, as models must also penalize false matches involving unseen identities.

Open-set evaluation frameworks therefore typically report:

• True Accept Rate (TAR)/False Accept Rate (FAR): The probability of correctly accepting a known identity versus incorrectly accepting an unseen identity at various thresholds.
• Detection and Identification Rate (DIR) vs. FAR: The fraction of true identifications among positive detections, plotted against the false accept rate.
• ROC curves quantifying the overall trade-off between sensitivity and specificity in open-set matching.

However, there is no mention in the literature of a specialized, principled "Genuine Open-set re-ID Metric" (GOM) being proposed as a distinct formalism or standard.

3. Latent Domain Expansion and Re-ID: Capacity and Evaluation

Recent advances in generative modeling and domain generalization focus on the latent structure and expansion capacity of feature spaces to accommodate previously unseen domains, including for tasks such as re-ID:

• Latent Domain Expansion (LDE) models explicitly construct latent spaces supporting multiple (possibly growing) domains, with orthogonality and disentanglement properties to mitigate feature collapse and destructive interference (Huang et al., 27 Jan 2026).
• LDE-style generators (e.g., in image generators) utilize dormant directions of latent space to encode new domains with minimal perturbation to existing representations, allowing expansion to hundreds of new domains without degrading performance on base domains (Nitzan et al., 2023).

While these works propose principled evaluation regimes—often reporting metrics such as mAP, accuracy, FID (for generative quality and domain preservation), and cluster separability—none define or standardize a "Genuine Open-set re-ID Metric" by this name.

4. Open-set Latent Domain Discovery and Recognition

Latent domain discovery methods for open-set or multi-source adaptation employ alignment strategies and side branches to partition samples into discovered domains with soft domain assignments (Mancini et al., 2018, Mancini et al., 2021). Feature normalization and domain alignment are performed per-discovered domain, with assignment quality and downstream recognition evaluated by:

• Performance recovery compared to oracle (domain-labeled) baselines.
• Cluster purity and interpretability (visual and quantitative clustering).
• Standard classification and retrieval accuracy under both closed- and open-set conditions.

No claim of a metric called GOM or "Genuine Open-set re-ID Metric" as a community standard or theoretical construct appears in these frameworks.

5. Discussion: Metric Gaps and Prospective Directions

The literature points to persistent challenges in open-set re-ID evaluation, including the need for principled metrics that:

• Separately quantify the model's ability to correctly match identities present in the gallery (genuine matches) and to robustly reject impostors (open-set capability).
• Balance domain expansion or adaptation capacity with the preservation of base domain discriminability.
• Remain robust to domain shift, scarcity of labeled identities, and label noise.

A plausible implication is that, while current work converges on broad open-set evaluation practices and principled latent-space engineering, the field lacks a universally acknowledged, theoretically grounded "Genuine Open-set re-ID Metric." The development and adoption of such a metric would require community consensus and rigorous benchmarking.

6. Related Literature

No attestation of GOM exists among the core references. However, the following are key contributions to latent domain expansion, domain discovery, and open-set recognition evaluation:

Paper Title	Contribution Scope	arXiv ID
Domain Expansion of Image Generators	Repurposing latent directions for new domain encoding	(Nitzan et al., 2023)
Boosting Domain Adaptation by Discovering Latent Domains	Side-branch domain assignment and VN-alignment	(Mancini et al., 2018)
Inferring Latent Domains for Unsupervised Deep DA	Joint latent domain discovery and classifier training	(Mancini et al., 2021)
Domain Expansion: A Latent Space Construction Framework	Orthogonal pooling for disentangled subspaces	(Huang et al., 27 Jan 2026)

A comprehensive survey of open-set re-ID metrics remains an open problem alongside the formalization of a genuine open-set re-identification metric.

7. Limitations and Outlook

Latent domain expansion, domain adaptation, and open-set recognition models continue to advance. However, no specific metric under the name "Genuine Open-set re-ID Metric (GOM)" has been defined or adopted in the current literature. The field instead utilizes a collection of open-set evaluation protocols drawn from verification, retrieval, and clustering diagnostics. Prospective research may formalize new metrics to robustly benchmark both the genuine-match identification and impostor rejection in open-set, multi-domain, and continually-expanding environments.