--- title: Genuine Open-set re-ID Metric (GOM) Overview url: https://www.emergentmind.com/topics/genuine-open-set-re-id-metric-gom type: topic --- # Genuine Open-set re-ID Metric (GOM) Overview A Genuine Open-set re-ID Metric (GOM) is not defined, mentioned, or implied in the factual record of the provided literature, including arXiv:2301.05225 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 [2601.20069]. - 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 [2301.05225]. 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 [1805.01386, 2103.13873]. 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| 2301.05225 | | Boosting Domain Adaptation by Discovering Latent Domains| Side-branch domain assignment and VN-alignment | 1805.01386 | | Inferring Latent Domains for Unsupervised Deep DA | Joint latent domain discovery and classifier training | 2103.13873 | | Domain Expansion: A Latent Space Construction Framework | Orthogonal pooling for disentangled subspaces | 2601.20069 | 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. Source: https://www.emergentmind.com/topics/genuine-open-set-re-id-metric-gom