AgroMap3D

AgroMap3D: Reliability-Aware Dense Mapping from Agricultural UAV Video

Javed AhmadShafiq AhmadDario AmiriFiora Pirri
DeepPlants, Rome, Italy
AgroMap3D combines local geometry and dual-rate reliability to assemble a dense field map from UAV video of repetitive crops, dry vegetation, and wind-driven vegetation
From agricultural UAV video to a dense field map.View figure ↗

From moving vegetation to coherent field maps.
Local geometry, temporal reliability, and field-scale pose refinement.

AgroMap3D dataset · synthetic-cue evaluation

22Agricultural sequences
42 minSelected evaluation windows
0.219 mMean aligned trajectory ATE
13.8%Lower mean ATE vs.
provider confidence alone

Abstract

Dense 3D mapping from agricultural UAV video is difficult due to repetitive crop structure, weak texture, thin vegetation, and wind-driven motion, which can make visual constraints ambiguous or temporally inconsistent. We present AgroMap3D, a reliability-aware framework that uses feed-forward 3D geometry as a local provider and reconciles relative constraints through pose-graph optimization. A dual-rate reliability mechanism evaluates stability on keyframe pairs and a denser temporal stream, down-weighting unreliable pose constraints while retaining vegetation geometry. When available, an uncertain global pose cue provides metric and field-scale context, and the refined poses assemble the local geometry into a dense map. We evaluate on our AgroMap3D dataset, comprising 22 agricultural UAV sequences, and on the public GrapeSLAM dataset using synthetically degraded global pose cues derived from trajectory references. Ablations on the AgroMap3D dataset show that local geometry reduces mean aligned trajectory error relative to the degraded cue, while reliability weighting further reduces mean absolute trajectory error relative to provider confidence alone. Qualitative reconstructions illustrate the effect of pose refinement on dense-map assembly.

See the reconstruction.

Each component contributes.

AgroMap3D dataset · Synthetic pose cue · Per-sequence Sim(3) alignment · Lower is better

Agricultural crop rows recorded in the AgroMap3D datasetAGROMAP3D DATASET

AgroMap3D Dataset

22 monocular RGB UAV sequences recorded with DJI Mavic 3 and DJI Mini platforms, paired with trajectory references extracted from onboard flight logs using AirData.

RGB videoTrajectory referencesEvaluation windows

88,784 source frames at 30 fps, with selected evaluation windows totaling 42.02 minutes. The collection spans repetitive crops, dry and sparse vegetation, range changes, long flights, turn-backs, revisits, and orbital motion.

Mavic 3 references use RTK-based positioning; Mini references use non-RTK positioning.

Dataset release Coming soon

BibTeX

ahmad2027agromap3d
@inproceedings{ahmad2027agromap3d,
  title     = {AgroMap3D: Reliability-Aware Dense
               Mapping from Agricultural UAV Video},
  author    = {Ahmad, Javed and Ahmad, Shafiq and
               Amiri, Dario and Pirri, Fiora},
  booktitle = {Proceedings of the IEEE/CVF Winter
               Conference on Applications of
               Computer Vision (WACV)},
  year      = {2027}
}