Temporal-Prior-Guided View Planning for Periodic 3D Plant Reconstruction

TL;DR

Periodic crop monitoring requires efficient 3D reconstruction across growth cycles; this paper introduces temporal-prior-guided view planning that reuses the previous-cycle plant model as a temporal prior. It aligns the prior to current partial data with a non-rigid embedded-deformation graph, inflates the reconstruction to anticipate growth, and solves a set-covering ILP to select a minimal, globally planned view set; a NBV step is included for robustness, and a global path minimizes movement. Experiments on maize and tomato show near-complete surface coverage with fewer views and comparable movement across hemisphere and sphere view spaces, outperforming NBV-only and prior one-shot baselines, with better generalization to plant complexity. The approach enables resource-efficient, periodic 3D plant reconstruction and can be extended with global motion planning and semantic cues for robustness.

Abstract

Periodic 3D reconstruction is essential for crop monitoring, but costly when each cycle restarts from scratch, wasting resources and ignoring information from previous captures. We propose temporal-prior-guided view planning for periodic plant reconstruction, in which a previously reconstructed model of the same plant is non-rigidly aligned to a new partial observation to form an approximation of the current geometry. To accommodate plant growth, we inflate this approximation and solve a set covering optimization problem to compute a minimal set of views. We integrated this method into a complete pipeline that acquires one additional next-best view before registration for robustness and then plans a globally shortest path to connect the planned set of views and outputs the best view sequence. Experiments on maize and tomato under hemisphere and sphere view spaces show that our system maintains or improves surface coverage while requiring fewer views and comparable movement cost compared to state-of-the-art baselines.

Figures (5)

• Figure 1: An example of the temporal prior construction for the target maize plant. The full 3D reconstruction from the previous cycle (red) and the partial observation acquired in the current cycle (blue) are registered to generate an approximated geometry, which serves as a temporal prior for subsequent view planning.
• Figure 2: Overview of the proposed view planning pipeline, following the combined framework in pan2024tro. The system starts from a random initial view to acquire an initial point cloud, which is then fused with a subsequent observation planned by the NBV module. Using temporal priors, it predicts the minimum set of views required to cover the plant to be reconstructed. Planned views are visualized with their local coordinate axes in red, green, and blue—while the initial view and NBV are indicated in black. Finally, the global path (purple) is computed to minimize the robot’s movement cost during data collection.
• Figure 3: Detailed pipeline of view planning with temporal priors. The framework first aligns the previously reconstructed 3D point cloud of the same plant (red)—serving as the temporal prior—with the current observation from the initial view and NBV (blue) via non-rigid registration. It then applies inflation (green) to account for potential plant growth. Finally, a set covering optimization problem is solved to determine the minimum view set required to fully cover the inflated geometry approximation.
• Figure 4: Examples of the two environmental configurations used in our experiments. (a) Hemisphere view space with a supporting table, shown with a Maize plant. (b) Sphere view space without a table, shown with a Tomato plant.
• Figure 5: Visual comparison of our method with GMC pan2023cviu and MA-SCVP pan2024tro. The top row shows the same initial view (black), the selected views (red-green-blue), and the corresponding view paths (purple). For a fair comparison, all methods are visualized with 12 views (our method and MA-SCVP both predict 12 views, while 12 views are selected from GMC for consistency). The bottom row visualizes the reconstructed plant surfaces (blue) and uncovered regions (red) using voxelization, focusing on the bottom part of the plant. Our method achieves the highest surface coverage with a shorter path; GMC also attains good coverage but with a longer path; MA-SCVP, despite a similar path length, fails to observe the bottom surface details.