GSAlign: Geometric & Semantic Alignment for AG-ReID

• The paper presents GSAlign, a transformer-based network that explicitly decouples geometric and semantic alignment to enhance aerial-ground person re-identification.
• It integrates a Learnable Thin Plate Spline module for spatial warping and a Dynamic Alignment Module for visibility-aware semantic filtering, achieving substantial performance gains on the CARGO benchmark.
• GSAlign outperforms conventional methods by significantly improving Rank-1 accuracy, mAP, and mINP, particularly in challenging cross-view matching scenarios.

Geometric and Semantic Alignment Network (GSAlign) is a transformer-based framework for aerial-ground person re-identification (AG-ReID) that explicitly addresses two distinct failure modes in cross-view matching: geometric misalignment induced by drastic viewpoint changes, and semantic misalignment induced by occlusion, partial visibility, and background clutter. Introduced for matching pedestrian images captured from unmanned aerial vehicles and ground-based surveillance cameras, GSAlign combines a Learnable Thin Plate Spline (LTPS) Module for geometric correction with a Dynamic Alignment Module (DAM) for visibility-aware semantic filtering, and reports substantial gains on the CARGO benchmark, particularly in the hardest aerial-to-ground retrieval setting (Li et al., 25 Oct 2025).

1. Task definition and problem formulation

AG-ReID aims to match images of the same pedestrian across drastically different viewpoints, typically between aerial/UAV cameras and ground-based surveillance cameras. In this setting, the same identity can appear structurally different across views, which makes AG-ReID more difficult than standard person re-identification. The paper isolates two core difficulties. The first is geometric misalignment, arising from strong pose distortion, changes in apparent body shape, scale variation, and rotation and perspective distortion. The second is semantic misalignment, arising from occlusions, partial visibility, background clutter, and different visible body regions across views (Li et al., 25 Oct 2025).

This decomposition is central to the design of GSAlign. Prior AG-ReID methods are described as mostly learning global shared embeddings or view-invariant representations, but often without explicit geometric correction or visibility-aware semantic filtering. GSAlign is therefore constructed around the claim that these two forms of mismatch should be modeled separately and explicitly.

A common misconception is to treat aerial-ground re-identification as merely a stronger domain adaptation problem. GSAlign instead frames it as a joint spatial-semantic alignment problem: local parts may not correspond under a simple spatial mapping, and the semantically informative evidence may differ across views even when the identity is the same. This suggests that cross-view person matching in AG-ReID requires more than a shared embedding space; it requires mechanisms that determine both where correspondence should be established and which evidence should be trusted.

2. Architectural organization

GSAlign is built on a ViT-Base backbone and follows the VDT architecture as its foundation. It introduces an additional view token and integrates two complementary modules: LTPS and DAM. The transformer tokenization is written as

$\left[ \mathbf{X}_{\text{cls}, \mathbf{X}_{\text{view}, \mathbf{X}_{\text{img} \right] = \text{VDT}\left( \left[ \mathbf{X}_{\text{cls}, \mathbf{X}_{\text{view}, \text{tokenize}(\mathbf{X}_{\text{img}) \right] \right),$

where $\mathbf{X}_{\text{cls}}$ denotes the class token, $\mathbf{X}_{\text{view}}$ the view token, and $\mathbf{X}_{\text{img}}$ the image tokens (Li et al., 25 Oct 2025).

The overall pipeline is organized as follows. The input image is tokenized and passed through a ViT/VDT-style transformer. LTPS is inserted progressively into transformer layers to perform geometric warping of patch features. In parallel, DAM operates during training to generate visibility-aware channel masks that filter semantic noise in class prototypes. The network is trained end-to-end with identity loss, triplet loss, deformation regularization, and mask alignment loss (Li et al., 25 Oct 2025).

The complementarity between the two modules is explicit in the paper. LTPS addresses where features should be aligned spatially, while DAM addresses which semantic dimensions should be trusted or suppressed. This division of labor gives GSAlign its defining character as a geometric-and-semantic alignment network rather than a conventional embedding model.

Component	Primary function	Operating level
LTPS	Geometric warping for cross-view spatial alignment	Patch-feature geometry
DAM	Visibility-aware masking for semantic filtering	Channel/prototype semantics
ViT-Base + VDT + view token	Backbone feature extraction and view conditioning	Transformer representation

Within the broader alignment literature, this architecture places GSAlign in a family of methods that avoid relying solely on latent invariance. Related work reinforces this orientation from other domains: GADL for graph alignment explicitly separates descriptor quality from latent-space geometric compatibility (Behmanesh et al., 11 Sep 2025), AlignGS uses semantics as an active geometric regularizer in sparse-view indoor reconstruction (Gao et al., 9 Oct 2025), and SGANet combines semantic and geometric alignment for multimodal multi-view anomaly detection (Bai et al., 7 Apr 2026). These parallels suggest a broader methodological pattern in which semantics and geometry are optimized jointly rather than sequentially or implicitly.

3. Learnable Thin Plate Spline module

LTPS is GSAlign’s geometric alignment mechanism. It extends classical TPS by making the transformation learnable and rotation-aware, and it is inserted into transformer layers in a progressive or hierarchical manner: shallow layers handle local deformation details, while deeper layers handle larger or more global pose variations (Li et al., 25 Oct 2025).

The module begins from a set of 2D source control points

$\mathbf{P}_s \in \mathbb{R}^{K \times 2},$

initialized as a regular grid over normalized coordinates $[-1,1] \times [-1,1]$ and treated as learnable parameters during training. A corresponding set of target control points

$\mathbf{P}_t \in \mathbb{R}^{K \times 2}$

is fixed and uses the same regular grid positions as the canonical target shape (Li et al., 25 Oct 2025).

A rotation angle is predicted from the current feature map,

$\theta = f_{\theta}(\mathbf{F}),$

and the source control points are rotated as

$$\mathbf{P}_s^{(\text{rot})} = \mathbf{P} \begin{bmatrix} \cos \theta & -\sin \theta \ \sin \theta & \cos \theta \end{bmatrix}^{\top}, \quad \theta = f_{\theta}(\mathbf{F}),$$

with

$$\theta \in \left[-\frac{\pi}{2}, \frac{\pi}{2}\right].$$

The TPS transformation satisfies the interpolation constraint

$\mathbf{X}_{\text{cls}}$0

and is written as

$\mathbf{X}_{\text{cls}}$1

where $\mathbf{X}_{\text{cls}}$2 and $\mathbf{X}_{\text{cls}}$3 (Li et al., 25 Oct 2025).

After warping, the LTPS output feature is fused residually with the original feature,

$\mathbf{X}_{\text{cls}}$4

where $\mathbf{X}_{\text{cls}}$5 is a tunable fusion factor. This preserves original semantic content while injecting explicit structural alignment. To avoid unstable or overly large rotations, LTPS is regularized by

$\mathbf{X}_{\text{cls}}$6

where $\mathbf{X}_{\text{cls}}$7 is the number of encoder layers with LTPS (Li et al., 25 Oct 2025).

Empirically, LTPS is the dominant contributor to cross-view gains. On CARGO, adding LTPS to the baseline changes A↔G performance from 48.12 / 42.76 / 29.95 to 64.89 / 61.08 / 50.54 in Rank-1 / mAP / mINP, indicating that explicit geometric correction is especially consequential when the viewpoint gap is large (Li et al., 25 Oct 2025).

4. Dynamic Alignment Module

DAM addresses semantic misalignment by predicting visibility-aware masks. Its motivation is that some regions in training images are occluded, irrelevant, or dominated by background noise, so the model should not treat all feature channels equally. DAM therefore dynamically adjusts semantic emphasis based on the current input (Li et al., 25 Oct 2025).

For each class $\mathbf{X}_{\text{cls}}$8, the batch prototype is computed as

$\mathbf{X}_{\text{cls}}$9

with prototype features $\mathbf{X}_{\text{view}}$0-normalized before and after update. The mask generator is a two-layer MLP followed by Sigmoid: $\mathbf{X}_{\text{view}}$1 where $\mathbf{X}_{\text{view}}$2, $\mathbf{X}_{\text{view}}$3, $\mathbf{X}_{\text{view}}$4, and $\mathbf{X}_{\text{view}}$5. The resulting channel-wise mask takes values in $\mathbf{X}_{\text{view}}$6 (Li et al., 25 Oct 2025).

The mask is applied to the class prototype via element-wise multiplication,

$\mathbf{X}_{\text{view}}$7

This suppresses irrelevant dimensions and emphasizes discriminative channels associated with visible body regions. DAM is used only during training; it affects prototype learning and alignment, adds no inference-time cost, and at inference features are compared with unmasked prototypes. The best-performing configuration is the Inner-Batch variant, where masks are generated from the current mini-batch and used to refine peer features (Li et al., 25 Oct 2025).

The associated mask loss is

$\mathbf{X}_{\text{view}}$8

where

$\mathbf{X}_{\text{view}}$9

and

$\mathbf{X}_{\text{img}}$0

The entropy term encourages decisive channel selection rather than trivial all-ones masks (Li et al., 25 Oct 2025).

Conceptually, DAM can be read as a semantic reliability estimator. Rather than aligning every feature coordinate indiscriminately, it filters the representation into a masked semantic subspace in which samples and class prototypes are encouraged to agree. This suggests that GSAlign’s semantic alignment is not merely cross-view invariance learning, but a selective process conditioned on visibility.

5. Optimization, data regime, and empirical performance

GSAlign is trained end-to-end with the total objective

$\mathbf{X}_{\text{img}}$1

where $\mathbf{X}_{\text{img}}$2 is the identity classification loss, $\mathbf{X}_{\text{img}}$3 is the triplet loss, $\mathbf{X}_{\text{img}}$4 regularizes LTPS, and $\mathbf{X}_{\text{img}}$5 governs DAM (Li et al., 25 Oct 2025).

The reported implementation settings are: ViT-Base backbone, ImageNet-21K pretraining, input size 256 × 128, horizontal flipping augmentation, AdamW optimizer, 0.05 weight decay, cosine learning-rate schedule, initial learning rate $\mathbf{X}_{\text{img}}$6, 20 epochs warm-up, 120 total epochs, batch size 64, and $\mathbf{X}_{\text{img}}$7 during training (Li et al., 25 Oct 2025).

Experiments are reported on three datasets: CARGO, AG-ReID, and AG-ReID v2. CARGO is the largest benchmark described in the paper, with 108,563 images, 5,000 identities, and 13 cameras comprising 5 aerial and 8 ground cameras. The train/test split is 51,451 images / 2,500 identities for training and 51,024 images / 2,500 identities for testing. Evaluation metrics are Rank-1 accuracy, mAP, and mINP, and on CARGO four protocols are used: ALL, G↔G, A↔A, and A↔G (Li et al., 25 Oct 2025).

The headline results are strongest on the most challenging cross-view setting. Under A↔G, GSAlign reaches 64.89% Rank-1, 61.55% mAP, and 52.81% mINP, compared to VDT with gains of +16.77% Rank-1, +18.79% mAP, and +22.86% mINP. Under ALL, GSAlign reports 65.06% Rank-1, 57.95% mAP, and 44.97% mINP, improving over VDT by +0.96 Rank-1, +2.75 mAP, and +3.84 mINP (Li et al., 25 Oct 2025).

The ablation study further clarifies module contributions. The baseline reports 64.10 / 55.20 / 41.13 on ALL and 48.12 / 42.76 / 29.95 on A↔G. Adding LTPS yields 64.42 / 55.95 / 41.92 on ALL and 64.89 / 61.08 / 50.54 on A↔G. Adding DAM on top of LTPS yields 65.06 / 57.95 / 44.97 on ALL and 64.89 / 61.55 / 52.81 on A↔G (Li et al., 25 Oct 2025). This pattern indicates that LTPS produces the primary cross-view improvement, while DAM mainly strengthens retrieval quality through better semantic filtering, especially as reflected in mAP and mINP.

6. Design choices, limitations, and relation to adjacent alignment frameworks

Several design ablations identify the operating regime of GSAlign. For LTPS, the paper tests 4, 9, 16, 25, 36 control points and finds that 4 control points provide the best overall trade-off, while more points can over-flex and slightly hurt mAP and mINP. For insertion depth, placements at the first layer, first 4 layers, middle 4 layers, last 4 layers, and all layers are compared, and all layers performs best, supporting the progressive alignment hypothesis. For DAM, Inner-Batch performs better overall than Memory Bank and Classification Matrix variants. Supplementary experiments further report that fixed-angle rotation is weaker, LTPS without $\mathbf{X}_{\text{img}}$8 is unstable, and LTPS with $\mathbf{X}_{\text{img}}$9 performs best (Li et al., 25 Oct 2025).

The paper also states several limitations. LTPS depends on accurate prediction of control point offsets and rotation angles; under extreme viewpoint changes or heavy occlusion, predicted transformations may be unreliable. In aerial-aerial settings, where geometric variation is smaller, LTPS and DAM may introduce unnecessary correction and can slightly degrade performance in some single-view scenarios. The method is therefore task-specific, designed primarily for aerial-ground settings with strong spatial misalignment (Li et al., 25 Oct 2025).

GSAlign should also be distinguished from other similarly named or conceptually adjacent alignment models. It is not the same as AlignGS, which targets sparse-view indoor reconstruction and treats semantic priors as a geometric regularizer in a 3D Gaussian representation (Gao et al., 9 Oct 2025). Nor is it the same as the graph alignment framework GADL, which combines dual-pass spectral encoding with geometry-aware functional maps for unsupervised graph correspondence (Behmanesh et al., 11 Sep 2025). Likewise, it differs from SGANet, which addresses multimodal multi-view anomaly detection through selective cross-view refinement, semantic-structural patch alignment, and explicit calibrated geometric correspondence (Bai et al., 7 Apr 2026). A plausible implication is that “geometric and semantic alignment” has become a cross-domain design principle rather than a single architecture family.

In that broader context, GSAlign is a domain-specific realization of a recurring methodological thesis: semantic invariance is insufficient when geometry is severely distorted, and geometric correction is insufficient when visible semantic evidence is incomplete. Its contribution lies in operationalizing this thesis for AG-ReID through a unified transformer with a learnable TPS-based warping mechanism and a training-time visibility-aware masking mechanism (Li et al., 25 Oct 2025).