VR3D: View-Robust 3D Representation Learning for Aerial-Ground Person Re-Identification

TL;DR

VR3D introduces a 3D-aware framework for aerial-ground person re-identification by mapping 2D image features into a canonical 3D space to reduce viewpoint bias. It uses 3D geometry-semantic attention to align body part features across views and a reliability-aware fusion mechanism to adaptively combine representations based on confidence scores.

Why It Matters

This approach improves cross-view person matching in real-world surveillance scenarios where aerial and ground cameras capture vastly different perspectives. By grounding 2D features in 3D space, it enhances robustness to occlusion and geometric distortion, enabling more accurate identification across platforms.

Key Findings

• VR3D achieves a 5.63% Rank-1 accuracy gain on the CARGO dataset compared to prior methods.
• The 3D geometry-semantic attention mechanism effectively aligns 2D patches with 3D voxels using spatial priors from single-view images.
• Reliability-aware fusion improves performance by dynamically weighting features based on sample-specific confidence, mitigating errors from poor 3D reconstructions.