acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Clustering and Label Weighting for Unsupervised Aerial-Ground Person Re-Identification

Abstract

Aerial-ground person re-identification (AG-ReID) aims to match images of the same person across aerial and ground views. To reduce manual annotation costs, recent studies have explored unsupervised AG-ReID (UAG-ReID). Existing UAG-ReID methods typically perform intra-view clustering followed by cross-view matching to obtain identity pseudo-labels. Although these methods reduce annotation costs, they overlook the reliability of pseudo-labels. Specifically, clustering algorithms inevitably introduce errors that can cascade and amplify during the matching process. To address these issues, we propose a dynamic clustering and label weighting method (DCLW) that introduces global dynamic clustering (GDC) and label dynamic weighting (LDW) to improve the reliability of clustering and mitigate the impact of noisy labels. Specifically, GDC jointly clusters aerial and ground images while dynamically adjusting the proportion of images from each view among the K nearest neighbors of each image. This avoids cascading errors caused by the two-stage process and aligns the results of clustering more closely with the prior data distribution. Based on GDC, LDW reweights image pseudo-labels to further mitigate the impact of noisy labels. In addition, DCLW introduces a mixture-pure contrastive loss which leverages single-view clusters to supplement cross-view supervision and promote view-invariant identity learning. We evaluate DCLW under multiple cross-view datasets, and the results demonstrate its superiority and the effectiveness of its components.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
aerial-ground person re-identification, unsupervised learning, clustering algorithm, contrastive learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_5LjfOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets unsupervised aerial-ground person ReID and argues that the common two-stage pipeline, intra-view clustering followed by cross-view matching, lets clustering noise propagate and become amplified. The proposed DCLW replaces this with global dynamic clustering, which jointly clusters aerial and ground features while allocating the K nearest neighbors according to the current aerial and ground image ratio. It then introduces label dynamic weighting to soften noisy pseudo-labels inside mixed clusters and a mixture-pure contrastive loss to exploit single-view pure clusters as extra cross-view supervision. Experiments on AG-ReID.v1 and LAGPeR show consistent gains over GGC and strong UVI-ReID baselines, with ablations supporting GDC, LDW, and the pure-cluster loss.

Critical strengths
  1. The paper identifies a concrete failure mode of existing UAG-ReID methods rather than only proposing another loss. The two-stage error-cascade story is directly supported by Figure 1 and the comparison among CVM, GCE, GDC, and intra-view clustering in Figure 3 and Figure 4. The dynamic neighbor allocation is tied to a real dataset prior.
  2. LDW and the mixture-pure loss are complementary and the ablations are convincing. LDW gives measurable gains over GDC alone, and L_pure further improves cross-view positives without hurting intra-view positives, which matches the distance-distribution analysis in Figure 6. The method does not discard pure clusters, turning them into additional cross-view supervision is a practical way to reduce wasted pseudo-labels.
Critical weaknesses

None

Questions for authors
  1. In GDC, k_a and k_g are computed from n_a and n_g, but it is not fully clear whether these counts include only clustered images, DBSCAN noise points, or all extracted features before clustering. If one view dominates or many points become noise, does the Jaccard/K-reciprocal structure remain stable, and could the dynamic ratio flip too aggressively between early and late training?
  2. LDW depends on the overlap between a GDC view cluster and the intra-view local cluster, while both are derived from the same evolving encoder. How sensitive is LDW to the quality of local clusters, and does the h-neighbor average risk confirming the same GDC errors when nearby samples are also mislabeled?
Reviewer_8n3SOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a clustering and label weighting method for unsupervised aerial-ground person re-identification (AG-ReID). The authors point out that existing clustering approaches are prone to noisy pseudo-labels, and they introduce a K-NN-based neighbor aggregation scheme together with a dynamic weighting mechanism that adjusts pseudo-label weights rather than weighting all samples equally. Experiments on AG-ReID benchmarks show that the proposed approach achieves better performance than existing state-of-the-art methods.

Critical strengths
  1. The overall pipeline and proposed approach are straightforward and easy to follow.

  2. The authors evaluate their method on multiple cross-view person re-identification datasets, including AG-ReID.v1 and LAGPeR, showing that more reliable pseudo-labels combined with the proposed model improve recognition performance.

Critical weaknesses
  1. How nearest neighbors are allocated in Global Dynamic Clustering (GDC) needs further mathematical and experimental justification. In Equation 4, k_a and k_g are updated using the counts of aerial and ground images (n_a and n_g) retained in local clusters. However, because the total ratio of aerial to ground images in a dataset is fixed, the proportion n_a / (n_a + n_g) should stay nearly unchanged across training epochs once most images are clustered. This implies that k_a and k_g may act like fixed constants rather than dynamic variables. To verify the actual contribution of this dynamic design, the authors could provide 1) a plot of how n_a / (n_a + n_g) and k_a, k_g change across epochs, and 2) an ablation comparing adaptive GDC with a static baseline where k_a and k_g are fixed to the overall dataset ratio.

  2. Using pure clusters in the mixture-pure contrastive loss (L_pure) carries a high risk of introducing false positive pairs. In Section 3.5.2, for each aerial pure cluster, the model picks the top m ground pure clusters with the highest centroid similarity as the positive set P_u^a. Since these pure clusters were not merged during global clustering by GDC, they likely belong to different people who only share similar appearance cues like clothing color or background. Pulling them together under L_pure with a soft target of 1/m can reinforce wrong cross-view matches in the feature space. The authors could evaluate the true identity matching accuracy of the top m selected pure clusters across epochs to show how much label noise this step introduces.

  3. Label Dynamic Weighting (LDW) depends on local clustering consistency to calculate the weight gamma in Equation 5. In early training epochs, however, the learned features are still noisy and unreliable, so both local and global clusters contain heavy noise. When the local neighbor labels y_(i,j) in Equation 6 are wrong, the weighted label hat_y_i mixes features from different identities and weakens feature discriminability. It would be helpful for the authors to clarify how their method avoids degrading feature learning through soft labeling in the early stages of training.

Questions for authors

Please see weakness. Also for some captioning like Table 1, please include the clear annotation for what UAG, SAG etc represents. It only appears in the first page which is away from where it is used.

2 comments

We sincerely thank the reviewer for the constructive suggestion, from which we have greatly benefited. We respond to your concerns as follows. The revisions in the manuscript are highlighted in blue font.

1. Regarding clarification of GDC. We apologize for not describing this clearly in the earlier version. We provide the following clarification: After the intra-view contrastive learning stage, GDC continues to cluster images separately within each view at every epoch, producing aerial local clusters and ground local clusters. Specifically, we use half of the original distance threshold to explore more reliable similarity relationships, and thereby estimate the number of nearest neighbors within each view based on these more reliable similarity relationships. A smaller distance threshold also prevents the vast majority of images from being grouped into clusters. GDC counts the aerial and ground images assigned to these local clusters, with the respective counts denoted by na{n_a} and ng{n_g}.

2. Regarding experiment of GDC. Following the reviewer's suggestion, we have added two additional experiments to the main paper. To further examine the behavior of GDC, we visualize the value of na/(na+ng)n_a/(n_a+n_g) over the entire training process. As shown in Figure 5, on both datasets this value changes dynamically throughout training. This is because we adopt a smaller distance threshold for the clustering that produces nan_a and ngn_g, so that neither count approaches the total number of training images. We further introduce an ablation variant, A4, which fixes nan_a and ngn_g to the total number of training images in the corresponding view. We find that A4 performs slightly worse than A3. For example, on AG-ReID.v1 (A$\to$G) and AG-ReID.v1 (G$\to$A), A3 outperforms A4 in Rank-1 by 2.12% and 2.25%, respectively; on LAGPeR (A$\to$G) and LAGPeR (G$\to$A), A3 outperforms A4 in Rank-1 by 2.13% and 0.30%, respectively. These results indicate that GDC can improve pseudo-label reliability and model performance by dynamically selecting K nearest neighbors according to the data distribution prior.

3. Regarding false positive pairs of mixture-pure contrastive loss. We greatly appreciate the reviewer's suggestion. We evaluated whether the mm pure clusters selected by the mixture-pure contrastive loss contain images with the same identity as the given cluster. Figure 8 shows the probability that the selected mm pure clusters contain images with the same identity as the given cluster throughout the training process. As shown in Figure 8, the probability that the selected mm pure clusters contain positive samples exceeds 60% throughout training, which is generally an acceptable probability for unsupervised tasks. These results indicate that Lpure{L_{pure}} can further promote view-invariant identity learning by leveraging pure clusters, thereby improving cross-view recognition performance.

4. Regarding LDW. We thank the reviewer. LDW is not used while features are unreliable: Stage 1 is a pure intra-view warm-up (Eqs. 1–3), and LDW only starts in Stage 2, with clustering re-estimated every epoch. Eq. 5 mixes labels, not features. A single wrong neighbor can contribute at most 1/h = 12.5% of its mass. L_intra also remains active in Stage 2 (Eq. 13), so within-view discriminability is never governed by the soft label. Empirically, adding LDW to GDC improves Rank-1 by 7.78/6.83 (AG-ReID.v1) and 4.67/3.31 (LAGPeR) under this exact schedule (Table 2), which would not happen if early soft labels degraded learning.

5. Regarding UAG, SAG etc represents. We thank the reviewer for the suggestion. In Section 4.2, we have added the following category descriptions: supervised aerial-ground ReID (SAG-ReID), unsupervised traditional ReID (UT-ReID), unsupervised visible-infrared ReID (UVI-ReID), and unsupervised aerial-ground ReID (UAG-ReID), to improve clarity.

Thank you for the reply, which addressed my concerns here. I have modified the rating accordingly.

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

This paper proposes Dynamic Clustering and Label Weighting (DCLW) for unsupervised aerial-ground person re-identification. The proposed method aims at addressing the problem of propagating noisy pseudo-labels during intra-view clustering and cross-view matching. Specifically, DCLW introduces Global Dynamic Clustering (GDC) to generate cross-view pseudo-labels by adjusting the numbers of aerial and ground neighbors according to the current sample distribution. Label Dynamic Weighting (LDW) then combines global pseudo-labels with neighboring pseudo-label information from intra-view clusters. In addition, a mixture-pure contrastive loss is introduced to exploit potential cross-view associations involving single-view clusters. Experiments on AG-ReID.v1 and LAGPeR demonstrate improvements over several existing unsupervised ReID methods. Ablation studies further indicate that the proposed components provide complementary performance gains.

Critical strengths
  1. Clear problem motivation and practical relevance: This paper addresses the problem of noisy pseudo-labels in unsupervised aerial-ground person re-identification, an important challenge for learning identity-discriminative representations without manual annotations. The motivation is supported by clustering visualizations and cross-view matching statistics.

  2. Proposed components with encouraging empirical results: The proposed GDC, LDW, and mixture-pure contrastive learning components target different aspects of pseudo-label generation and utilization. The ablation results show consistent performance improvements when these components are introduced, and the evaluation covers two aerial-ground ReID datasets.

Critical weaknesses
  1. Insufficient experimental comparisons and evaluation reliability.
  • Limited comparisons: GGC is the only baseline under the same UAG-ReID setting, and its results on LAGPeR are missing. The comparison with recent unsupervised cross-modal ReID methods could also be included to strengthen the experiments.
  • Inconsistent evaluation protocols: The reported AG-ReID.v1 A→G gallery size differs between DCLW (3,255) and the original GGC paper (3,331). It is unclear whether the reported results are directly comparable.
  1. Insufficient justification and analysis of the proposed components.
  • Limited justification of GDC: GDC adjusts the neighbor allocation using a dataset-level sample ratio. However, no comparison with a fixed allocation based on the aerial/ground data ratio is provided, making it difficult to determine whether dynamic updates are necessary.
  • Unverified reliability assumption in LDW: The overlap ratio between local and global clusters is used as a measure of pseudo-label reliability, but its correlation with actual pseudo-label correctness is not quantitatively validated.
  • Potential noise in mixture-pure contrastive learning: The selected cross-view positive clusters may contain different identities, while mixed-cluster centroids used as negatives may include false negatives. The reliability of these assignments is not sufficiently analyzed.
Questions for authors
  1. Could the authors clarify the discrepancy in the AG-ReID.v1 A→G gallery size between this submission and the original GGC paper? Were all results evaluated using exactly the same query and gallery splits? Could GGC also be evaluated on LAGPeR?
  2. How much of the improvement from GDC is attributable to dynamically updating the neighbor allocation ratio, rather than using a fixed ratio derived from the aerial/ground sample distribution? A controlled comparison would help establish the necessity of the dynamic design.
  3. Could the authors quantitatively analyze the relationship between the overlap weight used by LDW and the actual correctness of pseudo-labels? Additionally, is the weighted pseudo-label renormalized after removing components corresponding to same-view pure clusters?
  4. How reliable are the selected cross-view positive and negative clusters in the mixture-pure contrastive loss? How does the proposed method address potential false negatives among mixed-cluster centroids?
  5. Which value of h was used to produce the main experimental results: 6 or 8? Please clarify the inconsistency between the implementation details and the sensitivity analysis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.