ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Long-Sequence Strawberry Bed Reconstruction with VGGT Submaps

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600121.(doi:10.13031/aim.202600121)
Authors:   Ruoyao Qin, Mr., Won Suk Lee, Dr., Zijing Huang, Mr., Henry Medeiros, Dr.
Keywords:   3D Reconstruction, feed-forward multi-view geometry, long-sequence mapping.

Abstract. Large-scale 3D reconstruction of strawberry fields from ground imagery is needed to provide spatially explicit canopy geometry for plant-level phenotyping and to support precision agriculture tasks such as yield estimation and canopy volume assessment. However, bed-scale reconstruction remains challenging because acquisition trajectories are long and nearly linear, loop closure is typically unavailable, and canopy structure is highly repetitive. These conditions worsen drift and submap misalignment in conventional multi-view reconstruction workflows. In this study, a long-sequence reconstruction pipeline is presented that uses the Visual Geometry Grounded Transformer (VGGT), a feed-forward multi-view model that predicts camera parameters, dense depth and point maps, and long-range 3D point tracks, to generate locally consistent submaps from overlapping sequence batches. Submaps are stitched into a globally consistent bed-scale map by deriving inter-batch 3D to 3D correspondences from VGGT tracking, converted into stable view keys, and intersected across overlapping batches. Submap-to-submap motion is estimated under a motion-constrained model aligned with row traversal, using planar translation and roll-only rotation, improving robustness under noisy correspondences and inaccurate geometry. Global alignment is obtained through streamed pose chaining, followed by overlap-aware pruning and sliding window deduplication to suppress duplicated structures. An incremental voxel-hash fusion then integrates retained points into a unified global point cloud. The approach enables efficient reconstruction of complete strawberry beds from long RGB sequences, producing a continuous, large-scale 3D map suitable for downstream phenotyping and precision crop management.

(Download PDF)    (Export to EndNotes)