- 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
- 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.
- 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
- 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?
- 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?
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 and ng.
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) 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 na and ng, so that neither count approaches the total number of training images. We further introduce an ablation variant, A4, which fixes na and ng 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 m 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 m 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 m pure clusters contain positive samples exceeds 60% throughout training, which is generally an acceptable probability for unsupervised tasks. These results indicate that Lpure 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.