UAV3DCrop benchmark shows NeRF, 3DGS fail agronomic accuracy tests

Researchers from multiple institutions, including the University of Minnesota, introduced UAV3DCrop, a public benchmark designed to evaluate 3D reconstruction in repeated multi-angle UAV crop surveys. The dataset contains 88,830 RGB images at 5280x3956 pixels with a ground sampling distance of 3.6–5.8 mm, spanning 91 scenes of corn, soybean, wheat, and oat. Track A evaluates seven scene-optimized NeRF and 3D Gaussian Splatting variants on held-out views, photogrammetry-referenced depth, and canopy-height recovery. Track B tests four pretrained feed-forward models on zero-shot camera-pose and geometry estimation.

Results show that current reconstruction methods are not interchangeable for agronomic use: Splatfacto-big leads on appearance, while Scaffold-GS leads on depth and is statistically tied with Splatfacto for canopy height. No single method wins across all three targets. Among feed-forward models, MapAnything leads on seven of eight metrics, but others fail severely on absolute scale, a flaw that alignment conceals. Repeated acquisitions further reveal sensitivities to sequence position and tie-point multiplicity. The benchmark is publicly available, providing a rigorous testbed that could drive better 3D models for precision agriculture.