GeoUniPR: A Geometry-Consistent Unified Framework for Cross-Modal Place Recognition
Authors: Wonbong Kim, Jiatong Xiao, Rui Li, Xufei Wang, Qiwen Gu, Junqiao Zhao, Chen Ye, and Guang Chen
Submitted: 10 August 2026
Category: Computer Vision and Pattern Recognition (cs.CV)
Paper: arXiv:2608.11263 [cs.CV]
Abstract
Cross-modal place recognition (CMPR) aims to identify the same location across heterogeneous sensing modalities, such as vision and LiDAR. Existing methods typically bridge the modality gap using complex alignment modules, multi-stage training pipelines, or full fine-tuning of pretrained backbones. In this work, we revisit CMPR from the perspective of geometric consistency and propose GeoUniPR, a unified and concise geometry-consistent framework.
GeoUniPR reduces cross-modal discrepancy at the representation level by projecting LiDAR point clouds into the camera perspective to construct Geometry-Consistent Depth Image Views (DIV), which establish direct RGB–LiDAR correspondence. We further augment DIV with native LiDAR cues—including intensity and surface-normal information—yielding a multi-channel geometric representation that improves structural consistency. Based on this representation, GeoUniPR learns a unified embedding space using two modality-specific Vision Transformer (ViT) encoders with identical architectures. The model is trained through parameter-efficient adaptation, eliminating the need for auxiliary alignment modules, multi-stage training, or full backbone fine-tuning.
In addition, we introduce Spatially-Consistent InfoNCE (SC-InfoNCE), a CMPR-specific contrastive objective that suppresses distance-induced false negatives under spatial continuity—a common pitfall in prior contrastive approaches. Extensive experiments on the KITTI and KITTI-360 datasets demonstrate that GeoUniPR achieves state-of-the-art (SOTA) performance in both same-modal and cross-modal place recognition, with strong cross-dataset generalization.
1. Introduction
Place recognition is a fundamental capability for autonomous systems, enabling them to localize within previously visited environments. Cross-modal place recognition extends this task to scenarios where the query and reference data come from different sensor modalities, such as cameras and LiDAR. This is especially relevant in multi-robot systems and heterogeneous sensor fleets, where different platforms may rely on different sensing hardware.
Despite significant progress, existing CMPR methods often rely on heavy machinery: complex cross-modal alignment layers, multiple training stages, or fine-tuning large pretrained models. These approaches are computationally expensive and can be difficult to scale across diverse sensor configurations. Moreover, they often overlook the inherent geometric structure shared between modalities, which could serve as a powerful source of supervision.
In light of these limitations, we present GeoUniPR—a framework that leverages geometric consistency as the primary bridge between visual and LiDAR data. By projecting 3D point clouds into camera perspective and augmenting depth with native LiDAR cues, we create a representation that is both information-rich and naturally aligned with RGB images. Our approach simplifies the training pipeline while achieving superior performance, as evidenced by extensive experiments on public benchmarks.
2. Method Overview
GeoUniPR consists of three key components:
- Geometry-Consistent Depth Image Views (DIV): LiDAR point clouds are projected onto the camera image plane using known calibration parameters, producing dense depth maps. These depth views are spatially aligned with RGB images, establishing direct pixel-level correspondence.
- Multi-Channel Geometric Augmentation: To further enhance structural fidelity, we augment DIV with intensity and surface-normal channels derived from the raw LiDAR measurements. The resulting multi-channel representation captures both geometric and reflectance properties of the scene, improving robustness to appearance variations.
- Unified Embedding Learning: Two ViT-based encoders—one for RGB images and one for the multi-channel geometric representation—share identical architectures. They are trained jointly using a contrastive objective with our proposed SC-InfoNCE loss, which adapts standard InfoNCE to account for spatial continuity in continuous environments. This prevents the model from being penalized for matching nearby locations that are visually distinct but geographically close.
- State-of-the-Art Performance: GeoUniPR consistently outperforms existing methods across all tasks, setting new SOTA records on both datasets.
- Parameter Efficiency: Our parameter-efficient adaptation strategy achieves high accuracy while updating only a small fraction of the backbone parameters, reducing training cost and memory footprint.
- Cross-Dataset Generalization: Models trained on KITTI generalize effectively to KITTI-360 without additional fine-tuning, demonstrating the robustness of geometric consistency as a transferable signal.
3. Experiments and Results
We validated GeoUniPR on the KITTI and KITTI-360 datasets, which are widely used benchmarks for autonomous driving perception. Our evaluations cover both same-modal (RGB-to-RGB and LiDAR-to-LiDAR) and cross-modal (RGB-to-LiDAR and LiDAR-to-RGB) place recognition tasks.
Key findings include:
4. Conclusion
GeoUniPR demonstrates that geometric consistency can serve as a simple yet powerful unifying principle for cross-modal place recognition. By aligning representations through depth-based views and leveraging multi-channel LiDAR cues, we eliminate the need for complex alignment modules and multi-stage training. Our work highlights the potential of geometry-driven representation learning for robust, scalable, and efficient CMPR systems, and we believe it paves the way for further research into unified multi-modal perception frameworks.
For citation purposes, please refer to the original arXiv paper: arXiv:2608.11263.
via ArXiv CV