Problem Formulation. Standard AG-ReID (01) relies on paired data, which is often unavailable in real-world scenarios and leads to a missing cross-view challenge. To address this, we propose the Single-View Aerial-Ground Re-Identification (SV AG-ReID) setting (02), where models must adapt from a single viewpoint to an unseen target viewpoint.
Abstract
Aerial-Ground Re-Identification (AG-ReID) is constrained by the viewpoint-domain gap, as drastic viewpoint disparities occlude or distort discriminative features, making cross-viewpoint image retrieval challenging. While existing methods rely on paired cross-view annotations, real-world deployments, such as wilderness search-and-rescue (SAR), often lack target-domain data, requiring retrieval from ground-level references alone. To our knowledge, we are the first to address this challenge by formalizing the Single-View AG-ReID (SV AG-ReID) setting, where models trained on a single real viewpoint must generalize to an unseen viewpoint. We propose 3D Lifting-based Elevated Novel-view Synthesis (3D-LENS), a unified framework combining geometrically-consistent novel view synthesis that leverages large-scale 3D mesh reconstruction, with a robust representation learning scheme to mitigate synthetic-to-real bias. Unlike 2D generative baselines that suffer from geometric inconsistencies or prior 3D methods that are restricted to class-specific templates, our approach ensures view-consistent synthesis across diverse categories without predefined templates that fail to capture fine-grained details, such as carried objects. Extensive experiments demonstrate that our method achieves state-of-the-art performance on SV AG-ReID scenarios.
Proposed Solution. To tackle the SV AG-ReID, 3D-LENS uses a Geometrically Consistent Novel View Synthesis (01) to generate consistent novel views from single images using large-scale 3D synthesis, and a Robust Representation Learning scheme (02) to mitigate remaining biases in the synthetic data. This approach enables robust cross-viewpoint representation learning without requiring paired cross-view supervision.
Overview of the 3D-LENS framework. For the Geometrically Consistent Novel View Synthesis, a source image Ii is canonically lifted to a 3D representation (01), followed by novel view synthesis (02) and synthetic-to-real alignment (03) to produce target-view images Iisty. In our Robust Representation Learning scheme, an elevation-based curriculum scheduler (04) progressively introduces these synthetic views into training, then a balanced sampling strategy (05) combines the synthetic views Iisty and real images Ii to form a mixed batch, which is optimized using standard ReID losses and a domain loss (06).
Results
State-of-the-Art Comparison on AG-ReID Dataset. Training is performed on a single viewpoint (Aerial or Ground) and evaluated in a cross-view Query → Gallery scenario on Aerial (A), and Ground (G). Results are reported as mAP and Rank-1 (%). Best results are in bold, and second-best are underlined.
State-of-the-Art Comparison on AG-ReID.v2 Dataset. Training is performed on a single viewpoint (Aerial, CCTV or Wearable) and evaluated in a cross-view Query → Gallery scenario against Aerial (A), CCTV (C) and Wearable (W). Results are reported as mAP and Rank-1 (%). Best results are in bold, and second-best are underlined.
The margin separating 3D-LENS from prior arts across all cross-view scenarios reveals how severely conventional ReID methods degrade once they are placed in the SV AG-ReID setting. By leveraging large-scale 3D reconstruction for novel view synthesis, our method learns to bridge the viewpoint-domain gap without requiring any real data from the target viewpoint.
Qualitative Results
Qualitative results of geometrically consistent novel-view synthesis. For each example, we show the original input image alongside two synthesized novel views generated by our 3D lifting pipeline. The rendered views preserve identity-discriminative details such as clothing, body shape, and carried objects, while accurately reflecting the target viewpoint geometry.
Citation
@inproceedings{grolleau2026lens,
author = {Grolleau, William and Sabourin, Astrid and Lapouge, Guillaume and Achard, Catherine},
title = {{3D-LENS}: {A} {3D} Lifting-based Elevated Novel-view Synthesis method for Single-View Aerial-Ground Re-Identification},
booktitle = {European Conference on Computer Vision (ECCV)},
