Skip to content

Hierarchical Hyperbolic Representation Learning for Aerial-Ground Person Re-Identification

Conference: ECCV 2026
Paper: ECCV Official
Code: https://github.com/YangQiWei3/HiHR
Area: Remote Sensing
Keywords: Aerial-Ground Person Re-Identification, Hyperbolic Representation Learning, Hierarchical Metric Learning, Vision-Language Alignment, Lorentz Manifold

TL;DR

To tackle representation collapse and the suppression of view-specific details caused by forced Euclidean alignment under extreme aerial-ground viewpoint gaps, HiHR introduces Text-guided Multi-granularity Fusion (TMF) and Hierarchical Hyperbolic Learning (HHL) to preserve coarse identity separability and fine-grained view-specific cues in a Lorentz manifold, setting new state-of-the-art benchmarks on CARGO and LAGPeR.

Background & Motivation

Person Re-Identification (ReID) serves as a foundational capability for intelligent surveillance and multi-camera tracking. Conventional ReID setups primarily assume homogeneous ground-level camera networks where pedestrians are captured from roughly similar elevations and horizontal perspectives. However, real-world deployment increasingly demands the integration of heterogeneous camera platforms: unmanned aerial vehicles (UAVs) provide flexible, wide-area spatial coverage from oblique high-altitude viewpoints, while fixed ground cameras capture detailed close-range frontal and lateral views. This motivates Aerial-Ground Person Re-Identification (AG-ReID), which aims to match identities across aerial and ground sensors under severe perspective disparities and extreme scale variations.

Despite steady progress, existing AG-ReID approaches predominantly enforce direct cross-view feature alignment in Euclidean space, attempting to pull aerial and ground image embeddings of the same identity into a single tight cluster. This forced invariance introduces a fundamental flaw: it heavily penalizes viewpoint differences, treating view-specific discriminative cuesโ€”such as top-down head-shoulder contours and hairstyle patterns from UAVs, versus lower-body gait and footwear textures from ground camerasโ€”as unwanted variance to be suppressed. Consequently, these models suffer from representation collapse and lose critical local discriminability. Furthermore, most prevailing frameworks rely solely on high-level global semantic features extracted from the final layer of a backbone, attenuating mid-level structural cues and fine-grained patch patterns that are indispensable for distinguishing pedestrians with similar attire.

This paper tackles the challenge by reconsidering the geometric nature of cross-view identity representation: matching pedestrians across heterogeneous views inherently exhibits a tree-structured hierarchy, where a generic identity concept sits at the root while view-conditioned appearances form specialized leaves. Hyperbolic geometry, endowed with constant negative curvature and exponential volume expansion toward its boundary, offers an optimal inductive bias for accommodating such tree-like hierarchies without distortion. The core idea is the Hierarchical Hyperbolic Representation (HiHR) framework, which couples Text-guided Multi-granularity Fusion (TMF) with Hierarchical Hyperbolic Learning (HHL) on a Lorentz manifold to construct a two-level hierarchyโ€”a coarse parent level enforcing cross-view identity consistency and a fine child level preserving view-specific discriminative details via hyperbolic entailment cones.

Method

Overall Architecture

The HiHR pipeline comprises three interconnected stages: Multi-granularity Feature Extraction (MFE), Text-guided Multi-granularity Fusion (TMF), and Hierarchical Hyperbolic Learning (HHL). MFE extracts multi-layer visual tokens from a pre-trained CLIP vision encoder alongside dual-level text prompt embeddings. TMF leverages cross-attention to inject semantic guidance into Class tokens and adaptively aggregate Patch tokens across layers. Finally, HHL maps the fused features onto a Lorentz hyperbolic manifold with learnable scale separations and enforces fourfold entailment cone regularizations alongside dual-level metric supervision.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    IN["Input Image & Dual-Level Prompts"] --> MFE["Multi-granularity Feature Extraction<br/>ViT Multi-layer Tokens + Dual Prompts"]
    MFE --> TMF["Text-guided Multi-granularity Fusion<br/>Semantic Cross-Attention + Patch Aggregation"]
    TMF --> HHL["Hierarchical Hyperbolic Learning<br/>Lorentz Manifold Mapping + Entailment Cones"]
    HHL --> OUT["Logarithmic Mapping to Tangent Space<br/>Concatenated Dual-Level Retrieval Feature"]

Key Designs

1. Text-guided Multi-granularity Fusion: Semantic Cross-Attention and Patch Aggregation Relying exclusively on the final-layer Class token discards mid-level geometric cues and fine-grained local textures. To overcome this limitation, MFE extracts token sequences from intermediate CLIP image encoder layers \(\{5, 8, 12\}\). In parallel, dual-level prompts are constructed: a view-agnostic prompt (\(p_{vag}\): "A photo of [shared] person.") to capture generic identity evidence, and a view-aware prompt (\(p_{vaw}\): "A photo of [shared] person from [view-private] view.") to capture viewpoint-specific context. To bridge the vision-language modality gap, TMF employs the text embeddings as semantic queries to compute cross-attention over the concatenated multi-layer Class tokens \(\mathbf{C} = [\mathbf{c}^{(l_1)}; \dots; \mathbf{c}^{(l_M)}]\): $\(\boldsymbol{\alpha}_k = \mathrm{softmax}\left(\frac{\mathbf{Q}_k \mathbf{K}^\top}{\sqrt{D}}\right),\quad \mathbf{t}_k = \boldsymbol{\alpha}_k \mathbf{V},\quad k \in \{vag, vaw\}\)$ TMF reuses the attention weights \(\boldsymbol{\alpha}_k\) to dynamically aggregate multi-layer Patch tokens via \(\mathbf{X}_k = \sum_{m=1}^M \alpha_k^{(m)} \mathbf{X}^{(l_m)}\). A standard Transformer block subsequently fuses \(\mathbf{t}_k\) and \(\mathbf{X}_k\), yielding two complementary representations: the view-agnostic feature \(\mathbf{s}_{vag}\) (parent) and the view-aware feature \(\mathbf{s}_{vaw}\) (child).

2. Lorentz Hyperbolic Mapping and Scale Separation: Constructing the Geometric Hierarchy Directly optimizing parent and child representations in Euclidean space risks mutual interference. HHL adopts a Lorentz hyperbolic space with constant negative curvature \(-c\) (\(c > 0\)) as the embedding manifold. To initialize an explicit coarse-to-fine hierarchy, the normalized vectors are scaled in Euclidean space before projection: $\(\mathbf{g}^p = s_1 \frac{\mathbf{s}_{vag}}{\|\mathbf{s}_{vag}\|_2},\quad \mathbf{g}^c = s_1 s_2 \frac{\mathbf{s}_{vaw}}{\|\mathbf{s}_{vaw}\|_2},\quad s_1 > 0,\ s_2 > 1\)$ Here \(s_1\) regulates the overall embedding scale, while \(s_2 > 1\) positions child representations strictly farther from the origin than parent representations. Because hyperbolic space expands exponentially with geodesic distance from the origin, this geometric separation allocates substantially larger representational volume for fine-grained view-specific variations. Features \(\mathbf{g}^p\) and \(\mathbf{g}^c\) are then mapped to the Lorentz manifold via origin exponential mapping \(\exp_{\mathbf{o}}^c(\cdot)\), producing hyperbolic embeddings \(\mathbf{z}^p\) and \(\mathbf{z}^c\).

3. Fourfold Entailment Cone Regularization and Hierarchical Supervision: Preventing Cluster Drift While scale separation initializes radial depth, child clusters may still drift and overlap across different person identities. HHL computes Euclidean parent prototypes \(\boldsymbol{\pi}_y^p\) for each identity \(y\) within a mini-batch and projects them to hyperbolic prototypes \(\mathbf{z}^\pi = \exp_{\mathbf{o}}^c(\boldsymbol{\pi}_y^p)\). Each prototype defines an identity-specific entailment cone \(\omega(\mathbf{z}^\pi)\), within which the corresponding child embedding \(\mathbf{z}^c\) must reside: $\(\mathcal{L}_e(\mathbf{x}, \mathbf{y}) = \max\big(0, \phi(\mathbf{x}, \mathbf{y}) - \omega(\mathbf{x})\big)\)$ To enforce complete vision-language alignment across the hierarchy, HHL introduces four complementary entailment constraints: prototype-to-child visual containment \(\mathcal{L}_e(\mathbf{z}^\pi, \mathbf{z}^c)\), parent-to-child prompt containment \(\mathcal{L}_e(\mathbf{z}_P^p, \mathbf{z}_P^c)\), and same-level prompt-to-visual containment \(\mathcal{L}_e(\mathbf{z}_P^p, \mathbf{z}^p)\) and \(\mathcal{L}_e(\mathbf{z}_P^c, \mathbf{z}^c)\). Furthermore, supervision is decoupled: the parent feature \(\mathbf{g}^p\) is optimized via cross-view label-smoothed cross-entropy and triplet loss \(\mathcal{L}_g(\mathbf{g}^p)\) to enforce global cross-view compactness, whereas child features \(\mathbf{g}^c\) are supervised exclusively with intra-view triplet loss \(\mathcal{L}_{g,\mathrm{intra}}(\mathbf{g}^c)\) using anchors and candidates sampled from the identical view, ensuring view-specific cues remain unsuppressed.

Loss & Training

The complete multi-task optimization objective is formulated as: $\(\mathcal{L} = \mathcal{L}_g(\mathbf{g}^p) + \mathcal{L}_{g,\mathrm{intra}}(\mathbf{g}^c) + \lambda_e \Big[ \mathcal{L}_e(\mathbf{z}^\pi, \mathbf{z}^c) + \mathcal{L}_e(\mathbf{z}_P^p, \mathbf{z}_P^c) + \mathcal{L}_e(\mathbf{z}_P^p, \mathbf{z}^p) + \mathcal{L}_e(\mathbf{z}_P^c, \mathbf{z}^c) \Big]\)$ The model employs CLIP-ViT-B/16 with input images resized to \(256 \times 128\). Training runs for 60 epochs using Adam on an NVIDIA A800 GPU (batch size 64, sampling 4 images per identity). Initial learning rates are set to \(3.5 \times 10^{-4}\) for newly added modules and \(5 \times 10^{-6}\) for pre-trained weights, scheduled by cosine annealing after a 100-iteration warmup. Default hyperparameters are \(c=1.0, \lambda_e=5, s_1=0.5, s_2=2.0\). During inference, hyperbolic embeddings \(\mathbf{z}^p\) and \(\mathbf{z}^c\) are mapped back to tangent Euclidean space via logarithmic mapping \(\log_{\mathbf{y}}^c(\cdot)\), scaled by \(1/s_1\) and \(1/(s_1 s_2)\) respectively, and concatenated into a unified feature vector for cosine distance evaluation.

Key Experimental Results

Main Results

The authors evaluate HiHR across four AG-ReID benchmarks: AG-ReID, AG-ReID v2, CARGO, and the real-world captured LAGPeR dataset.

Table 2 from Original Paper: Performance Comparison on CARGO Benchmark (%)

Method ALL mAP ALL R1 Aโ†’G mAP Aโ†’G R1 Aโ†’A mAP Aโ†’A R1 Gโ†’G mAP Gโ†’G R1
VDT [33] 55.20 64.10 42.76 48.12 66.83 82.50 71.59 82.14
SeCap [22] 56.89 64.72 46.37 48.75 66.90 82.50 75.24 82.54
GSAlign [13] 57.95 65.06 61.55 64.89 65.55 80.00 73.86 83.04
SAS-VPReID [31] 58.48 64.42 57.36 58.51 63.41 70.00 76.90 83.93
LATex [32] 67.09 76.96 58.88 66.87 69.06 80.00 79.90 90.18
HiHR (Ours) 67.52 74.36 70.53 75.53 69.60 75.00 80.93 87.50

On CARGO under the challenging Aerial-to-Ground (Aโ†’G) protocol, HiHR achieves 70.53% mAP and 75.53% Rank-1, outperforming the previous state-of-the-art LATex (58.88% mAP) by an impressive +11.65% mAP and +8.66% Rank-1. On the real-world LAGPeR dataset (Table 3), HiHR leads with 33.80% mAP and 46.95% Rank-1 on Aโ†’G matching.

Ablation Study

Ablations conducted on the CARGO dataset validate the necessity of each architectural component and loss formulation.

Table 4 from Original Paper: Module Ablation on CARGO

Config ALL mAP ALL R1 Aโ†’G mAP Aโ†’G R1 Note
Baseline (CLIP-ViT) 63.61 68.23 66.06 66.02 Euclidean metric baseline
+ TMF 65.46 72.08 69.50 70.28 Text-guided multi-granularity fusion (+3.44% Aโ†’G mAP)
+ TMF + HHL (Full Model) 67.52 74.36 70.53 75.53 Full hierarchical hyperbolic framework (+1.03% mAP, +5.25% R1 on Aโ†’G)

Table 5 from Original Paper: Ablation of Entailment Losses on CARGO

\(\mathcal{L}_e(\mathbf{z}^\pi, \mathbf{z}^c)\) \(\mathcal{L}_e(\mathbf{z}_P^p, \mathbf{z}_P^c)\) \(\mathcal{L}_e(\mathbf{z}_P^p, \mathbf{z}^p) + \mathcal{L}_e(\mathbf{z}_P^c, \mathbf{z}^c)\) ALL mAP ALL R1 Aโ†’G mAP Aโ†’G R1
\(\times\) \(\times\) \(\times\) 64.62 71.15 68.06 72.34
\(\checkmark\) \(\times\) \(\times\) 65.46 72.12 69.50 70.21
\(\checkmark\) \(\checkmark\) \(\times\) 65.71 72.44 68.21 71.28
\(\checkmark\) \(\checkmark\) \(\checkmark\) 67.52 74.36 70.53 75.53

Key Findings

  • Dominant Gains on Ranking Precision (mAP): Across benchmarks, HiHR demonstrates the largest relative margins in mAP rather than solely Rank-1. This confirms that the hierarchical hyperbolic structure prevents false positive collapse and preserves fine-grained similarity rankings across the entire gallery.
  • Cross-Layer Complementarity: Evaluating ViT layers shows that \(\{5, 8, 12\}\) outperforms deep adjacent stacking \(\{10, 11, 12\}\). Intermediate layers provide vital structural and spatial cues that compensate for high-level semantic over-abstraction.
  • Importance of Quadruple Entailment: Omitting all entailment regularizations causes ALL mAP to fall from 67.52% to 64.62%, verifying that geometric containment is indispensable to prevent child cluster drift and maintain vision-language semantic alignment.

Highlights & Insights

  • Formulating Cross-View Retrieval as a Tree Hierarchy in Hyperbolic Space: Rather than treating viewpoint discrepancies as nuisance domain shifts to be forcibly eliminated, HiHR reconceptualizes the problem as an identity-to-view tree and leverages the exponential capacity of negative curvature to host view-specific details without identity confusion.
  • Dual-Level Prompt and Vision Co-Hierarchy: Aligning generic vs. view-private prompts with parent vs. child visual features creates a symmetric multimodal hierarchy, ensuring that text guidance directs the model to isolate true invariant traits from viewpoint-conditioned textures.
  • Transferable Metric Learning Paradigm: The combination of radial scale separation, entailment cone bounds, and decoupled cross-view/intra-view triplet supervision offers a general recipe for heterogeneous retrieval problems such as visible-infrared Re-ID and cross-view vehicle tracking.

Limitations & Future Work

  • Author-Acknowledged Computational Overhead: Calculating Riemannian exponential/logarithmic maps, along with cross-attention over multi-layer tokens and multiple entailment angles, increases training latency and GPU memory requirements compared to vanilla Euclidean models.
  • Inference Dimensionality: Retrieval features are formed by concatenating projected parent and child vectors, which doubles feature dimensionality and expands memory footprints during large-scale gallery indexing.
  • Future Directions: Exploring native hyperbolic approximate nearest neighbor (ANN) indexing to eliminate the need for tangent-space mapping during inference, and scaling to multi-UAV swarm surveillance.
  • vs VDT [33] / SeCap [22]: Prior methods attempted view-decoupling within flat Euclidean space, where limited capacity forces a trade-off between cross-view compactness and view-specific richness; HiHR bypasses this capacity bottleneck via hyperbolic negative curvature.
  • vs CLIP-ReID [14] / LATex [32]: While CLIP-based predecessors relied on single-level text prompts or late visual adaptation, HiHR develops Dual-Level Prompts coupled with multi-layer patch aggregation, enabling fine-grained semantic injection.

Rating

  • Novelty: โญโญโญโญโญ [Pioneering introduction of hierarchical hyperbolic learning and entailment cones to aerial-ground person ReID]
  • Experimental Thoroughness: โญโญโญโญโญ [Exhaustive evaluation across four benchmarks with granular layer selection, prompt design, and loss component ablations]
  • Writing Quality: โญโญโญโญโญ [Clear motivation, rigorous mathematical grounding, and well-structured diagrams]
  • Value: โญโญโญโญโ˜† [Offers an insightful geometric framework for cross-platform retrieval tasks under extreme viewpoint variations]