UAV3DCrop: Benchmarking 3D Reconstruction in Repeated Multi-Angle UAV Crop Surveys

## Abstract Accurate 3D crop monitoring is essential for data-driven precision agriculture, enabling field-scale analysis of plant structure, growth dynamics, and management responses. While modern 3D reconstruction methods perform well on generic benchmarks, their rendered appearance does not always translate into metrically and agronomically useful geometry in real crop fields. To address this gap, we introduce UAV3DCrop, a public benchmark comprising repeated multi-angle unmanned aerial vehicle (UAV) surveys of agricultural fields. The dataset includes 88,830 RGB images at 5280 × 3956 pixels resolution, with a ground sampling distance of 3.6–5.8 mm, captured across 91 scenes covering corn, soybean, wheat, and oat. The benchmark is organized into two tracks. Track A evaluates seven scene-optimized methods—including Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS) variants—on held-out views, photogrammetry-referenced depth, and canopy-height recovery. Track B tests four pre-trained feed-forward models on zero-shot camera-pose and geometry estimation. Our results reveal that scene-optimized methods rank differently across the three targets: Splatfacto-big leads in appearance, while Scaffold-GS excels in depth and is statistically tied with Splatfacto for canopy height. Among feed-forward models, MapAnything outperforms others on seven of eight metrics, whereas the remaining models show greater variability across crops and exhibit severe failures in absolute scale, issues that alignment can conceal. Repeated acquisitions further expose sensitivities that vary by output type and model, linked to position within the acquisition sequence and tie-point multiplicity. These findings indicate that current 3D reconstruction methods are not yet interchangeable for agronomic use: no single method excels simultaneously in appearance, geometry, and canopy height, and only one of four feed-forward models recovers usable metric scale. The dataset is publicly available at [https://link-dev.github.io/UAV3DCrop/](https://link-dev.github.io/UAV3DCrop/).

via ArXiv CV