acceptodds
Under review as a conference paper at ICLR 2027

GUIDE: Geometry-Grounded Identity Structure Learning for Multi-Modal Object Re-Identification

Abstract

Multi-modal object re-identification (ReID) aims to retrieve the same target across cameras by exploiting complementary cues from heterogeneous modalities. Recent graph-based approaches have further advanced multi-modal ReID by explicitly modeling relational dependencies among local regions and semantic representations. However, graph-based methods typically derive relational topology from feature or semantic similarity, causing corresponding regions across modalities to aggregate different contextual information during relational propagation, thereby weakening the consistency of identity-related structural cues. To address this issue, we propose GUIDE, a Geometry-Grounded Identity Structure Learning framework for multi-modal object ReID. GUIDE decouples relational topology from modality-dependent appearance by grounding connectivity in patch geometry while preserving modality-specific representations. Specifically, the Grid Graph Encoder (GGE) restores patch tokens to their native two-dimensional layouts and propagates local identity cues over geometry-defined neighborhoods using a graph encoder shared across modalities. The Axial Strip Readout (ASR) preserves spatial composition by aggregating graph-refined tokens into ordered strips with learnable queries. Together, GGE and ASR produce structured identity representations that complement global semantic features for retrieval. Extensive experiments on RGBNT201, RGBNT100, and MSVR310 demonstrate the effectiveness of GUIDE across person and vehicle ReID benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Object Re-identification, Multi-Modal Learning, Graph Neural Networks

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_h4QROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GUIDE for multi-modal object re-identification using RGB, near-infrared, and thermal images. Its central idea is to construct graph connectivity from fixed patch geometry rather than modality-dependent feature similarity. A Grid Graph Encoder performs local propagation, while an Axial Strip Readout aggregates the refined tokens into ordered strip representations. These structural features complement global ViT features, with auxiliary losses encouraging global–structural consistency, cross-modal consistency, and strip diversity. Experiments cover three person and vehicle benchmarks, including missing-modality evaluation.

Strengths

  • The framework is straightforward and achieves competitive performance across all three benchmarks. The component ablations on RGBNT201 show improvements from adding GGE, ASR, and the auxiliary losses, supporting the effectiveness of the overall design.

Weaknesses

  • The implementation details are incomplete. Several hyperparameters, including the number of graph layers, the number of strips, and the strip diversity margin, are not explicitly specified. The paper also does not analyze sensitivity to these choices, making it difficult to reproduce the results and assess their robustness.
  • Moreover, all quantitative component ablations are restricted to RGBNT201, leaving their contributions on vehicle datasets unclear. The choice of strip orientation is not experimentally examined, and whether spatial augmentations are synchronized across modalities is unspecified despite its relevance to the spatial-consistency motivation.
Critical strengths

The method is straightforward and improves retrieval performance across three benchmarks. The ablation results on RGBNT201 further support the effectiveness of its components, making it a relatively simple approach with demonstrated empirical benefits.

Critical weaknesses
  1. Incomplete implementation details and limited ablation coverage. Several hyperparameters, including the number of graph layers, the number of strips, and the diversity margin, are not explicitly specified. Although the loss weights are examined, sensitivity to these other choices is not reported. Furthermore, all component ablations are conducted on RGBNT201, leaving their effectiveness on the vehicle datasets unclear. These limitations hinder reproducibility and assessment of robustness.

  2. Limited methodological novelty and insufficient validation of key design choices. GGE uses fixed eight-neighbor aggregation with feature transformations and residual connections. What advantages does this design offer over conventional methods, and what quantitative evidence supports these advantages? Have alternative strip partitioning strategies been investigated? Is partitioning along the longer spatial dimension equally suitable for both pedestrians and vehicles? The experiments show differences in retrieval performance, but provide limited direct evidence that the proposed mechanisms work as intended.

Questions for authors
  1. Implementation details: Please specify the number of GGE layers, the number of strips, the diversity margin, and the inference fusion weight. Beyond the loss-weight analysis in Table 3, how sensitive are the results to these choices?

  2. Contribution of GGE: What advantages does fixed eight-neighbor aggregation offer over conventional local aggregation methods? Beyond the retrieval results in Table 4, is there quantitative evidence supporting the claimed improvement in cross-modal structural consistency?

  3. Strip partitioning: Have alternative strip orientations or partitioning strategies been evaluated? Is partitioning along the longer spatial dimension equally suitable for pedestrians and vehicles? Please also provide quantitative results for the ASR-only variant shown in the appendix.

  4. Ablation coverage: All quantitative component ablations are conducted on RGBNT201. Do GGE, ASR, and the auxiliary losses provide consistent benefits on the vehicle datasets?

  5. Data augmentation: Are random flipping and cropping synchronized across modalities, and are identical augmentation settings used for all ablation variants? Has the effect of these augmentations been examined, particularly given the method’s reliance on spatial organization?

  6. Feature dimensions and loss computation: Please clarify the output dimension of the ASR projection and how the global and structural representations are combined for LGLL_{GL}. Figure 2 lists this loss but does not clearly show its computation path.

  7. The qualitative visualizations focus on pedestrians. Why are vehicle examples not included, and can you provide them to assess whether the proposed structural modeling also works for vehicles?

Reviewer_VqVmOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies multi-modal object re-identification from a geometry-grounded structural learning perspective. The main idea is to decouple relational topology from modality-dependent appearance, allowing different modalities to share a common geometric structure while preserving their distinct visual representations. Based on this, the authors propose GUIDE, which first uses a Grid Graph Encoder (GGE) to propagate local identity cues through geometry-defined patch neighborhoods, and then employs an Axial Strip Readout (ASR) to aggregate graph-refined features into spatially ordered structural representations. Experiments on RGBNT201, RGBNT100, and MSVR310 show strong retrieval performance, together with additional analyses on component effectiveness, missing-modality robustness, and structural representation learning.

Critical strengths
  1. The paper presents a clear, well-motivated research problem, highlighting how existing graph-based multi-modal ReID methods rely on modality-dependent feature similarities to construct relational topology.
  2. GUIDE adopts a structured modular design, combining the Grid Graph Encoder (GGE) and Axial Strip Readout (ASR) to capture local relational information while preserving spatial structure.
  3. The pipeline is clearly presented, and motivation figures effectively distinguish similarity-based graph construction from the proposed geometry-grounded approach, enhancing readability.
  4. Experiments on three multi-modal ReID benchmarks, including component ablations, missing-modality evaluations, and qualitative visualizations, provide relatively comprehensive empirical evidence of GUIDE's effectiveness.
Critical weaknesses
  1. The motivation for decoupling relational topology from modality-dependent features is reasonable, but the advantages over similarity-based graph construction could be demonstrated more clearly. Although the appendix compares GGE with depth-wise convolution, a direct comparison using similarity-based graph topology would make the central motivation more convincing.
  2. The Related Work section could benefit from a clearer thematic organization. Separating representation learning and graph-based approaches into distinct subsections would improve the logical structure and readability.
  3. GUIDE achieves the highest mAP on RGBNT100, but its Rank-1 accuracy is lower than DeMo. Some discussion of this performance difference would help clarify the strengths and limitations of the proposed approach.
  4. The paper does not provide a systematic analysis of computational efficiency. Since GGE and ASR introduce additional graph propagation and query-guided aggregation, reporting inference time, FLOPs, and parameters would help clarify the efficiency–performance trade-off.
Questions for authors
  1. In Eq. (21), the explicit mathematical formulations of the cross-entropy and triplet losses are not provided, and Table 3 does not examine their individual contributions. Could the authors clarify their definitions and provide additional ablation experiments to evaluate their individual and combined effects on retrieval performance?
  2. The paper does not report inference time or provide a systematic analysis of computational efficiency. Since GUIDE introduces graph message passing in GGE and query-guided aggregation in ASR, could the authors provide comparisons of inference time, FLOPs, and parameters with existing methods to demonstrate the efficiency–performance trade-off?
  3. Although GUIDE achieves the highest mAP on RGBNT100, its Rank-1 accuracy remains lower than DeMo. Could the authors explain the possible reasons for this performance difference?
Reviewer_HtdLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the inconsistency of relational neighborhoods induced by modality-dependent feature similarity, the authors propose GUIDE for multi-modal object re-identification. Based on a shared CLIP-pretrained ViT, the framework introduces a Grid Graph Encoder (GGE) that propagates patch features over geometry-defined neighborhoods, and an Axial Strip Readout (ASR) that aggregates refined tokens into ordered strips using shared learnable queries. Global–structural consistency, strip diversity, and cross-modal consistency losses further regularize the structural representations, which are combined with global features for retrieval. Experiments on RGBNT201, RGBNT100, and MSVR310 report improvements in mAP, with additional evaluation under missing-modality conditions on RGBNT201.

Critical strengths
  1. GGE explicitly separates graph connectivity from modality-dependent patch features, while ASR preserves ordered spatial information without requiring semantic part annotations. The two modules provide complementary structural modeling mechanisms.
  2. GUIDE achieves the highest reported mAP on all three benchmarks. The RGBNT201 ablation also shows an improvement from 76.0% to 83.0% mAP, supporting the practical effectiveness of the complete framework.
Critical weaknesses
  1. The paper attributes the improvements to topology decoupling, but Tables 3–4 compare entire modules or different operators. A more targeted comparison would change only geometric versus feature-similarity adjacency while keeping propagation, readout, and losses unchanged.
  2. Table 3 reports the baseline, baseline+GGE, and baseline+GGE+ASR, whereas ASR alone appears only in the appendix visualizations. Its retrieval scores should also be reported to quantify GGE’s additional contribution beyond ordered strip aggregation.
  3. MDReID is included in Table 1 and discussed as supporting flexible multi-modal retrieval, yet it is absent from Table 2. The missing-modality comparison should include this method under the same evaluation protocol.
  4. GUIDE exceeds MGRNet by 0.3 and 0.5 mAP points on RGBNT100 and MSVR310, respectively. These small margins should be accompanied by means and standard deviations across independent runs to distinguish consistent gains from random variation.
  5. Section 3 introduces graph depth LL, strip count KK, diversity margin mm, and weighted feature fusion at inference. Their numerical settings are not provided in the implementation section or appendix.
Questions for authors
  1. How does GUIDE perform when geometric adjacency is replaced with feature-similarity adjacency under otherwise identical settings?
  2. What are the quantitative results for ASR alone, and how is the DWConv replacement matched to GGE in depth and capacity?
  3. How does GUIDE compare with MDReID under the same missing-modality protocol?

Comments on the paper

3 comments

Thanks for addressing my concerns on this paper. My concers have been solved.

Please remember to include the parameter details and analyses in the revised version.

I have improve my rating form 2 to 3.

Good luck.

Thanks for the authors' detailed answers! This has clarified some of my doubts. If the authors could further elaborate on the "Critical Weaknesses", it might enhance the overall quality of the paper even more.

Thanks to the authors for their structured responses! This has clarified my doubts. I have changed the score to positive.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.