Re-Ranking-Only Global Features

• Re-Ranking-Only Global Features are methods that refine retrieval by reordering fixed candidate lists using global descriptors and precomputed distances.
• They apply techniques such as graph propagation, MDS embedding, and neural language models to enhance performance in person re-identification, document retrieval, and visual place recognition.
• Empirical studies show significant gains in mAP and ranking metrics while preserving a fixed first-stage representation and modular deployment.

Re-Ranking-Only Global Features denotes, across the literature reviewed here, retrieval and recommendation pipelines in which the first-stage representation or candidate generator is held fixed and ranking quality is improved entirely by post-processing global descriptors, global scores, or neighborhood relations derived from them. In person re-identification, this often means operating only on Euclidean distances or rank lists built from a single embedding (Sarfraz et al., 2017); in document retrieval, it means injecting corpus-level document interactions after dense scoring (Francesco et al., 2024); in multi-view instance retrieval, privileged information may shape training while deployment remains restricted to global multi-view descriptors (Li et al., 2018); and in domain-generalizable Re-ID, an auxiliary MLLM can refine inference-time distances without changing the encoder (Li et al., 15 Jun 2026). The unifying systems constraint is that the re-ranking stage acts as a post-hoc module over already available global representations.

1. Definitional core and scope

In the surveyed work, a method belongs to this category when the inference-time interface is deliberately narrow: the module receives a fixed shortlist, a fixed global feature vector, a fixed score, or a fixed distance matrix, and returns only a reordered list. The representation extractor is not retrained during re-ranking, and the candidate set is usually unchanged. This is explicit in ECN, where “these rank lists and the distance matrix are the only inputs,” and in the music re-ranking setting, where the method “does not change which items are recommended, only their order” (Sarfraz et al., 2017); (Gong et al., 2020).

The phrase “global features” is used in several technically distinct but compatible senses. In person Re-ID and image retrieval, it typically denotes a single whole-image descriptor $\mathbf{x} \in \mathbb{R}^d$, as in the 1,536-dimensional embedding of PSE or the 2,048-dimensional and 8,448-dimensional descriptors used in VPR (Sarfraz et al., 2017); (Hanning et al., 15 Apr 2025). In recommendation, the same idea appears as whole-item feature vectors or whole-list representations built from existing ranking features, rather than part-wise or token-wise alignment (Pei et al., 2019). In local-to-global retrieval, “re-ranking-only global features” can even be computed on-the-fly from local similarities and then used solely for second-stage ordering (Aiger et al., 4 Sep 2025).

This suggests that the category is defined operationally rather than architecturally. A method may be unsupervised and distance-based, graph-based, prompt-based, or learned with auxiliary supervision; what remains constant is that deployment-time improvement is obtained after initial retrieval, using only global-level information available to the re-ranking stage.

2. Distance-matrix and rank-list formulations

A canonical formulation is the Expanded Cross Neighborhood (ECN) method. Given a probe image $p$, a gallery $G=\{g_i\mid i=1,\dots,B\}$, global embedding vectors, and squared Euclidean distance,

$$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$

one first computes all pairwise distances and the initial rank lists

$$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$

ECN then defines an expanded neighborhood from the top-$t$ immediate neighbors and the top-$q$ neighbors of each of those neighbors. For a probe,

$N(p, t) = \{ g_i^o \mid i = 1, 2, \dots, t \},$

$N(t, q) = \{ N(g_1^o, q), \dots, N(g_t^o, q) \},$

and the expanded neighbor multiset is

$N(p, M) \leftarrow \{ N(p, t), N(t, q) \},$

with

$p$0

The ECN distance is then

$p$1

The defining property is that no new descriptors, parts, or pair-specific rank lists are created; the procedure reads only top segments of already computed lists and averages pre-existing distances (Sarfraz et al., 2017).

ECN also admits a rank-list-based distance. Using the Jarvis and Patrick similarity,

$p$2

with $p$3, one obtains a normalized similarity $p$4 and then

$p$5

This can replace $p$6 inside the ECN formula, yielding a second purely post-hoc variant. The paper characterizes the whole procedure as unsupervised because no identity labels are used in the re-ranking stage, and automatic because once distances and rank lists are available, the procedure is deterministic (Sarfraz et al., 2017).

The same paper makes the computational contrast explicit. ECN “never creates reciprocal neighbor sets, does not compute Jaccard distances, and does not re-sort any list conditioned on another image,” whereas many contemporary re-rankers required pairwise reciprocal-neighbor recomputation. On Duke, the reported timings were 124.6 seconds for k-reciprocal re-ranking, 115.3 seconds for ECN (rank-dist), and 73.2 seconds for ECN (orig-dist), all on 19,889 images (Sarfraz et al., 2017).

3. From neighbor heuristics to graph-based global context

A major line of development reformulates re-ranking as message passing on graphs derived from global descriptors. One explicit statement is that re-ranking can be decomposed into “retrieving high-quality gallery samples” and “updating features,” where the first phase is building the $p$7-nearest-neighbor graph and the second phase is spreading messages within that graph (Zhang et al., 2020). In this view, for global features $p$8, one defines cosine similarity

$p$9

a $G=\{g_i\mid i=1,\dots,B\}$0-NN adjacency

$G=\{g_i\mid i=1,\dots,B\}$1

and a symmetric adjacency

$G=\{g_i\mid i=1,\dots,B\}$2

The node feature can be taken as the row of $G=\{g_i\mid i=1,\dots,B\}$3,

$G=\{g_i\mid i=1,\dots,B\}$4

and message passing becomes

$G=\{g_i\mid i=1,\dots,B\}$5

Under this interpretation, classical AQE, $G=\{g_i\mid i=1,\dots,B\}$6-QE, SCA, and $G=\{g_i\mid i=1,\dots,B\}$7-reciprocal re-ranking become special cases of graph propagation over global descriptors (Zhang et al., 2020).

The same work emphasizes the practical consequence of a sparse graph formulation. On Market-1501, the reported processing time is reduced from 89.2s for conventional k-reciprocal re-ranking to 9.4ms with one K40m GPU, while achieving mAP 94.65% and R@1 96.11% from a baseline of mAP 88.26% and R@1 95.28% (Zhang et al., 2020).

Document retrieval extends the same principle from feature-space neighborhoods to a corpus graph. GNRR builds a global semantic corpus graph

$G=\{g_i\mid i=1,\dots,B\}$8

over the full corpus, then restricts it per query to a query-induced subgraph whose node set is $G=\{g_i\mid i=1,\dots,B\}$9. Node features are based on the element-wise product $$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$0, optionally augmented with a BM25-related feature, and a GNN branch computes

$$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$1

These graph-conditioned features are concatenated with the independent branch and fed to a scorer. On TREC-DL19, the paper reports a relative improvement of 5.8% in Average Precision over the TCT baseline, from 0.430 to 0.455 (Francesco et al., 2024).

Visual place recognition adds another variant through affinity vectors. GCSA starts from global descriptors $$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$2, learns a projection $$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$3, defines visual affinity vectors

$$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$4

and concatenates them with optional affinity vectors from positional, heading, or radio side information,

$$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$5

These are projected and processed by a self-attention GNN. In the visual-only regime, the method still functions as a global-descriptor re-ranker; with side information, it generalizes descriptor-only contextual similarity aggregation into a heterogeneous affinity-space re-ranker. On Mapillary SLS with NetVLAD, the reported mAP@10 rises from 19.7 without re-ranking to 39.7 with GCSA, and Rec@10 from 50.7 to 75.1 (Hanning et al., 15 Apr 2025).

4. Training-time augmentation with global-only deployment

A second pattern keeps deployment-time re-ranking global-only while allowing richer supervision during training. DMVPIR is a direct example. Training uses both original global multi-view features $$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$6 and privileged ROI features $$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$7, learning original-space latent codes $$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$8, privileged-space latent codes $$d_{\text{orig}(p, g_i)} = \lVert \mathbf{p} - \mathbf{g}_i \rVert_2^2,$$9, view-specific matrices $$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$0, and dual hyperplanes $$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$1. The original-space multi-view reconstruction term is

$$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$2

and the SVM+-style coupling is

$$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$3

At test time, privileged features are unavailable and unused; only the original global multi-view descriptors are projected: $$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$4 and ranked by

$$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$5

With TE$$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$6 as initial ranking, the reported Oxford5k mAP rises from 62.04% to 81.51%, and Paris6k from 61.76% to 77.83% (Li et al., 2018).

MUSE follows the same deployment logic in DG person Re-ID. The encoder remains fixed and produces raw Euclidean distances $$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$7. A fine-tuned MLLM then predicts a pairwise distance from “Yes/No” logits: $$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$8 These are fused only for top-$$\mathcal{L}(p, G) = \{g_1^o, g_2^o, \dots, g_B^o\}.$$9 candidates: $t$0 The resulting $t$1 can be passed unchanged into K-RNN, ECN, or CAJ. On the MA$t$2C3 protocol, Baseline + K-RNN reports mAP 51.8 and Rank-1 45.9, while Baseline + MUSE + K-RNN reports 56.8 and 51.7; on MS$t$3C3, the corresponding numbers are 49.5/44.9 and 56.1/51.4 (Li et al., 15 Jun 2026).

A broader implication is that “global-only at test time” does not imply “global-only during learning.” Some methods use privileged information, hard-pair mining, or language-conditioned supervision to shape the re-ranking metric, yet preserve a strictly global, post-hoc inference interface.

5. Constructed and semantic global rerankers

A more radical variant constructs new global features specifically for re-ranking. In the local-to-global paradigm, initial retrieval is performed with efficient local feature search, then a query-specific $t$4 dissimilarity matrix $t$5 is assembled for the query and its top-$t$6 local neighbors. If a database–database pair is absent from the precomputed sparse local distances, the distance is set to $t$7. MDS is then applied to find an embedding whose Euclidean distances approximate these local dissimilarities, conceptually minimizing a stress of the form

$t$8

These MDS coordinates are then treated as “global features” and fused with SuperGlobal features for re-ranking. On $t$9Oxf+1M Hard, the reported mAP is 79.8 for the local-to-global system versus 77.7 for AMES; on $q$0Par+1M Hard, 83.4 versus 82.0 (Aiger et al., 4 Sep 2025).

MLLM-based semantic rerankers push the same idea into listwise reasoning. CoTRR assumes a fixed global-feature first stage, using CLIP / OpenCLIP or OSrCIR to produce a top-$q$1 candidate set, then performs three prompt-driven stages: query deconstruction,

$q$2

candidate evaluation,

$q$3

and listwise ranking,

$q$4

The backbone retrieval model is unchanged; the MLLM reasons over full images and full query semantics as global entities. On Flickr30K with CLIP ViT-B/32, the reported R@1 improves from 66.56 for base CLIP to 84.68 for CoTRR, and on MSCOCO from 39.45 to 58.31 (Wu et al., 18 Sep 2025).

These methods indicate that the “global feature” abstraction can be semantic rather than purely geometric. A plausible implication is that once the first stage supplies a sufficiently good shortlist, re-ranking can be framed either as metric refinement in descriptor space or as global reasoning over whole-item representations.

6. Empirical profile, misconceptions, and limitations

Across the surveyed papers, the main empirical pattern is consistent: keeping the candidate set or encoder fixed still leaves substantial headroom for improvement. On Market-1501 with IDE features, ECN (rank-dist) improves mAP from 55.0 to 71.1 and Rank-1 from 78.9 to 82.3; on MARS it improves mAP from 41.2 to 53.2 (Sarfraz et al., 2017). In document retrieval, GNRR yields AP 0.455 versus 0.430 for TCT on TREC-DL19 (Francesco et al., 2024). In VPR, GCSA raises LaMAR validation mAP@10 from 44.6 without re-ranking to 61.2 with visual+radio+positional affinities (Hanning et al., 15 Apr 2025). In DG Re-ID, MUSE turns conventional re-ranking into a stronger inference-time module without modifying the encoder (Li et al., 15 Jun 2026). These results make clear that first-stage global descriptors often contain recoverable structure not fully exploited by direct nearest-neighbor sorting.

A compact view of representative outcomes is given below.

Method	Fixed component at deployment	Reported outcome
ECN (Sarfraz et al., 2017)	Global embeddings and distance matrix	Market-1501 IDE mAP 55.0 $q$5 71.1
GNN re-ranking (Zhang et al., 2020)	Global descriptors	Market-1501 processing 89.2s $q$6 9.4ms
GNRR (Francesco et al., 2024)	BM25 top-1000 and TCT embeddings	TREC-DL19 AP 0.430 $q$7 0.455
GCSA (Hanning et al., 15 Apr 2025)	Global descriptors, optional side info	MSLS NetVLAD mAP@10 19.7 $q$8 39.7
MUSE (Li et al., 15 Jun 2026)	Re-ID encoder	MA$q$9C3 K-RNN mAP 51.8 $N(p, t) = \{ g_i^o \mid i = 1, 2, \dots, t \},$0 56.8
L2G (Aiger et al., 4 Sep 2025)	Local shortlist; on-the-fly global embedding	$N(p, t) = \{ g_i^o \mid i = 1, 2, \dots, t \},$1Oxf+1M Hard 79.8 vs 77.7 for AMES

Several misconceptions recur. One is that re-ranking-only methods can compensate for arbitrarily weak first-stage features. Multiple papers state the opposite in different forms: ECN assumes nearest neighbors are meaningful in the initial embedding, CoTRR can only re-rank the top-$N(p, t) = \{ g_i^o \mid i = 1, 2, \dots, t \},$2 candidates it is given, and candidate-set post-filtering in music recommendation leaves the item set unchanged (Sarfraz et al., 2017); (Wu et al., 18 Sep 2025); (Gong et al., 2020). Another is that “global-only” implies no auxiliary information. DMVPIR uses privileged ROI features during training, GCSA can exploit positional or radio affinities, and MUSE uses an MLLM-derived pairwise signal, yet all three preserve a global-only deployment interface for re-ranking (Li et al., 2018); (Hanning et al., 15 Apr 2025); (Li et al., 15 Jun 2026).

The main limitations are equally consistent. Complexity often remains quadratic in the number of compared items because full or partial distance matrices, graph construction, or shortlist-wise reasoning are still required (Sarfraz et al., 2017); (Zhang et al., 2020). Performance depends on shortlist quality, since re-ranking typically cannot recover items absent from the candidate set (Gong et al., 2020); (Francesco et al., 2024); (Wu et al., 18 Sep 2025). Methods that rely on auxiliary supervision or side information introduce annotation, engineering, or inference costs, whether through ROI boxes, corpus graphs, MLLM calls, or sensor metadata (Li et al., 2018); (Francesco et al., 2024); (Li et al., 15 Jun 2026); (Hanning et al., 15 Apr 2025). This suggests that the central design trade-off in Re-Ranking-Only Global Features is not whether post-hoc refinement works—it frequently does—but how much structure, supervision, and computation can be added while preserving the modularity that defines the paradigm.