Loop-Free Long-Sequence Field Mapping with VGGT

A VGGT-based 3D mapping pipeline with an overlap-guided submap strategy for stable, drift-free reconstruction of large agricultural fields without loop closure.

Long-sequence field mapping is challenging when loop closure is unavailable or unreliable. In this project, we leverage a Visual Geometry Grounded Transformer (VGGT) as the backbone for 3D reconstruction and introduce an overlap-guided submap iterative reconstruction strategy to achieve stable, drift-controlled mapping in large-scale agricultural environments.

VGGT reconstruction overlay on a full strawberry field — Overview of the reconstructed strawberry block captured without loop closure.

Zoom-in of submap alignment — Submap overlap keeps long tractor rows aligned through each sweep.

VGGT detail around irrigation fixtures — Irrigation fixtures retain sharp structure after iterative refinements.

Overlap-guided stitching view — Overlap-guided stitching blends boundary regions without visible seams.

Row centerline reconstruction detail — Row centerlines stay coherent over repeated tractor runs.

Methodology

The method decomposes long trajectories into manageable, overlapping segments:

VGGT-based reconstruction. Use VGGT to estimate geometry and camera poses along the sequence, taking advantage of its strong global reasoning capabilities in structured field scenes.
Overlap-guided submaps. Partition the sequence into submaps with controlled overlap, reducing the accumulation of drift within each segment.
Iterative alignment. Align and refine submaps iteratively using overlap regions as constraints, propagating corrections while keeping the overall structure consistent.
Drift-aware evaluation. Quantify drift and stability across long distances to validate the mapping performance compared with naive single-pass approaches.

Results & Impact

The overlap-guided VGGT pipeline produces stable, drift-free 3D reconstructions of large fields even without explicit loop closure, supporting long-range analytics such as spatial variability mapping and route planning.

This work demonstrates how modern transformer-based geometry backbones can be combined with careful sequence structuring to handle real-world, large-scale agricultural mapping tasks.