PaperlayerView exact revision on arXiv

Source-linked AI summary

Denoising-Aware Temporal Point Cloud Completion for 3D Crop Architecture Recovery and Phenotypic Trait Extraction

Mrudul Mittal, Soumyashree Kar

arXiv:2608.28343v1cs.CV

TL;DR

Temporal crop reconstruction lacks complete geometric ground truth and remains under-explored for agricultural architecture. The paper introduces SynthCrop4D and a denoising-aware temporal completion pipeline, finding that domain-specific denoising improves reconstruction and trait extraction.

  • Problem

    Agricultural 3D phenotyping lacks complete geometric ground truth and denoising-temporal completion pipelines that exploit longitudinal structure.

  • Method

    The paper introduces SynthCrop4D and evaluates denoising combined with temporal point cloud completion using previous-stage information on synthetic and real crop data.

  • Results

    Denoising improves reconstruction quality, with the best reported configuration achieving a Chamfer Distance of 0.0061 on SynthCrop4D and an F-Score of 0.2080 on Pheno4D.

  • Takeaways & Limitations

    The benchmark and pipeline support temporal plant reconstruction and more accurate phenotypic trait extraction across synthetic and real crop data.

  • Takeaways & Limitations

    Maize height fitting is weak and other maize traits lack meaningful sigmoid trajectories because Pheno4D covers the mid-vegetative V5–V9 stage.

Abstract

from arXiv · show

High-throughput phenotyping depends on accurate 3D reconstruction of plants across growth stages, yet the development and evaluation of temporal completion methods are limited by the lack of datasets with complete geometric ground truth. To address this challenge, we introduce SynthCrop4D, a procedurally generated synthetic dataset of temporally evolving plant point clouds that provides controllable noise, occlusion, and complete plant geometry for benchmarking reconstruction methods. Using this dataset, we evaluate a two-stage pipeline that combines spatial denoising and temporal point cloud completion. First, a denoising module removes structural artifacts from raw laser-scanned point clouds. The resulting data are then processed by an Adaptive Temporal PoinTr model that reconstructs the current growth stage (t) using information from the previous stage (t-1), enabling recovery of regions missing due to self-occlusion. We evaluate the proposed framework on both SynthCrop4D and the real-world Pheno4D dataset (tomato and maize) under settings with and without denoising. Results show that denoising substantially improves reconstruction quality, with the best configuration achieving a Chamfer Distance of 0.0061 on SynthCrop4D (Temporal PoinTr + Mamba-DG) and an F-Score of 0.2080 on Pheno4D (Vanilla PoinTr + Mamba-DG). We further demonstrate the use of completed point clouds for phenotypic trait extraction, including plant height, canopy width, and convex hull volume, obtaining hull-volume MAEs of 0.021 on synthetic data and 0.343 on real data. Together, SynthCrop4D and the proposed pipeline provide a benchmark and methodology for temporal plant reconstruction and high-throughput crop phenotyping.

1 Introduction

The paper targets crop point-cloud reconstruction under sensor noise and self-occlusion, introducing a synthetic benchmark and a denoising-aware temporal completion pipeline. Its evaluations show that preprocessing and temporal modeling support reconstruction and phenotypic trait extraction.

  • Sensor noise distorts thin crop structures, while self-occlusion creates missing leaf surfaces that hinder direct phenotyping measurements.
  • Existing methods lack crop-specific denoising biases, discard temporal information, and rarely evaluate denoising and completion as coupled tasks.
  • SynthCrop4D provides 20 maize and tomato plants across six stages with paired clean/noisy point clouds and controllable occlusion for supervised evaluation.
  • Mamba-DG uses a Hilbert-curve serialized selective state-space encoder to aggregate global context with O(n) complexity for crop denoising.
  • Adaptive Temporal PoinTr fuses the current partial scan with the t−1 growth stage through temporal cross-attention, whereas a dense-decoder temporal baseline reaches Test CD = 0.287.

2 Related Work

Related work shows that existing denoising and completion methods largely process point clouds independently and were not designed for agricultural noise, temporal structure, or trait-aware validation. The paper addresses these gaps with Mamba-DG, Adaptive Temporal PoinTr, and a systematic cross-domain ablation.

  • PointCleanNet performs efficient local displacement regression, while score-based and graph-based methods provide alternative denoising strategies for point clouds.
  • PCN, PoinTr, and Point-MAE establish major completion paradigms, but these methods generally treat scans independently.
  • Pheno4D supplies longitudinal tomato and maize scans, yet temporally aware crop reconstruction remains under-explored relative to single-timepoint reconstruction.
  • Agricultural denoising methods must address structured-light artifacts and canopy occlusions, while downstream denoiser effects on reconstruction remain underinvestigated.
  • The paper proposes Mamba-DG with Hilbert-serialized state-space modeling and Adaptive Temporal PoinTr with cross-attention, evaluated in a 12-configuration ablation.

3.1 Problem Formulation

The formulation represents the current crop scan as noisy and partially occluded, pairs it with the previous growth-stage scan, and evaluates completion with or without denoising.

  • The current-stage input is a noisy, partially occluded crop point cloud with a 30% radial crop simulating self-occlusion.
  • The previous-stage input is the SOR-filtered scan of the same plant and serves as the temporal prior.
  • Case 1 performs reconstruction without denoising, using the identity function as the preprocessing stage.
  • Case 2 applies one of four denoising models before one of three completion architectures.

3.2 Datasets and Preprocessing

The study combines real Pheno4D scans with procedurally generated SynthCrop4D sequences, using preprocessing and plant-level splits to evaluate reconstruction under distinct noise and ground-truth conditions.

  • Pheno4D Dataset: Pheno4D contains laser-scanned tomato and maize plants captured across 5–6 temporal growth stages.
  • Pheno4D Dataset: Real scans lack exact clean geometry, and registration noise is entangled with self-occlusion, preventing controlled isolation of the two degradations.
  • SynthCrop4D Dataset: SynthCrop4D procedurally generates maize and tomato geometry with explicit ground truth, enabling controlled noise, occlusion, and reconstruction evaluation.
  • SynthCrop4D Dataset: The synthetic dataset contains 20 plants and six growth stages, split by plant index into 14 training, 3 validation, and 3 testing plants.
  • SynthCrop4D Dataset: Synthetic growth scales height, leaf length, and organ counts linearly from 33% to 100%, deliberately differing from real maize’s sigmoidal trajectory.
  • Preprocessing: Synthetic clouds receive isotropic Gaussian noise with σ = 0.01, while real scans undergo outlier removal, voxel downsampling, and normalization to [−1, 1].

3.3 Stage 1: Denoising Pipeline

The denoising stage compares displacement-, graph-, and score-based architectures, including Mamba-DG, which combines Hilbert serialization with global state-space processing and score-based iterative refinement.

  • Baseline denoisers: PCN uses a PointNet-style shared-MLP encoder with three Conv1D layers, global max pooling, and per-point displacement regression.Its training loss is CD-L1.
  • Baseline denoisers: GCN builds a k-NN graph and applies EdgeConv-style convolutions, recomputing neighborhoods in feature space after each layer.Local edge features are aggregated by max pooling.
  • Evaluation design: The study compares four denoisers and three completion architectures across Pheno4D and SynthCrop4D.The dataset comparison distinguishes real-data limitations from synthetic controlled geometry.
  • Score-based denoiser: PDN augments PointNet-derived per-point features with an Umbrella Operator Feature and trains a ScoreNet using denoising score matching.Inference uses Langevin dynamics for 10 steps with step size 0.2.
  • Mamba-DG: Mamba-DG replaces PDN’s local PointNet encoder with Hilbert-serialized Mamba SSM blocks for O(n) global context aggregation.Features return to 3D order before local k-NN fusion, ScoreNet prediction, and Langevin denoising.

3.4 Stage 2: Completion Architectures

The completion stage ranges from a displacement-only PCN baseline to temporal fusion models that use prior growth-stage geometry to reconstruct missing regions.

  • PCN baseline: PCN encodes the partial scan into a 256-dimensional descriptor and decodes a fixed-size point set, but cannot add points outside the partial scan’s support.It is included as a lower-bound reference because its displacement-only design limits recovery of absent geometry.
  • Dense temporal fusion: The dense-decoder temporal baseline concatenates global descriptors from the current partial and prior clean frames before fully connected decoding.Despite regularisation and BatchNorm handling, it collapsed and underperformed a non-temporal self-attention baseline.
  • Adaptive Temporal PoinTr: Adaptive Temporal PoinTr tokenizes current and prior scans into proxy points via FPS and encodes them with a shared Conv1D encoder.The current features are spatially refined before temporal fusion.
  • Adaptive Temporal PoinTr: A Transformer decoder uses the spatially refined current features as Query and prior-stage features as Key/Value to query earlier geometry.A coarse MLP followed by a fine Conv1D head generates proxy completions and local point patches.
  • Ablation: Vanilla PoinTr is an ablation that removes prior-stage features and the cross-attention decoder.This isolates the contribution of temporal conditioning.

3.5 Mesh-Derived Trait Extraction

The pipeline evaluates biologically relevant structure by converting point clouds into meshes and extracting primary and secondary geometric traits.

  • Mesh reconstruction: Ground-truth and reconstructed point clouds are converted to meshes with Poisson surface reconstruction after normal estimation and tangent-plane orientation.Surface area is extracted as a leaf-area proxy, alongside a binary watertightness flag.
  • Trait descriptors: Secondary descriptors include vertical center of mass, projected canopy area, effective LAI proxy, canopy closure, compactness ratio, and elongation ratio.These are supplementary descriptors rather than primary evaluation metrics because several are derived ratios.

3.6 Delta-Conditioning and Trait-Aware Fine-Tuning

The V2 pipeline addresses temporal growth mismatch and trait-level errors through delta-conditioned prior features and trait-aware fine-tuning on real data.

  • Motivation: Baseline Adaptive Temporal PoinTr overestimated hull volume by importing prior-stage geometry without accounting for interstage plant growth.Its CD-L2 loss could also remain low while macroscopic traits were inaccurate.
  • Delta-conditioning: Delta-conditioning estimates growth from current and prior proxy tokens and uses an adaptive gate to weight the correction per proxy.The growth-adjusted prior replaces the static prior as Key/Value in temporal cross-attention.
  • Trait-aware fine-tuning: Trait-aware fine-tuning augments CD-L2 with a trait consistency term penalising discrepancies in predicted height and canopy width.The terms are computed directly from output points, with λh = λw = 0.1 and negligible training overhead.
  • Architecture: The Adaptive Temporal PoinTr architecture uses current partial and complete prior scans, spatial refinement, temporal cross-attention, and coarse-to-fine decoding.This architecture is illustrated as the temporal completion design underlying the pipeline.

4 Experimental Setup

The experiments compare denoising and completion pipelines across two evaluation regimes, using standardized geometric, overlap, runtime, trait, and temporal-growth metrics. Models retain architecture-specific denoising objectives while completion models share a common training recipe.

  • Metrics: Chamfer Distance measures bidirectional nearest-neighbor error, with lower values indicating better reconstruction.Results use CD-L1 or CD-L2 depending on the evaluation section.
  • Metrics: F-Score @ 5mm measures the harmonic mean of precision and recall within a normalized threshold, with higher values indicating better overlap.The metric is reported alongside CD because reproducing input points can inflate it when scans are not denoised.
  • Metrics: Trait MAE, relative trait MAE, and growth-delta sign agreement evaluate phenotypic accuracy and temporal trend consistency.Relative MAE supports comparison across domains with different trait scales, while sign agreement is scale-invariant.
  • Evaluation design: The study evaluates 12 denoising–completion combinations across raw and denoised inputs.Case 1 uses raw noisy scans; Case 2 applies one of four denoisers before one of three completion models.
  • Temporal setup: Temporal inputs consist of a partial current scan and the clean scan of the same plant from the previous growth stage.The partial scan is normalized with the ground truth and temporal prior using its center and scale; the initial stage has no prior.
  • Training: Denoisers use architecture-specific objectives, whereas all completion networks share AdamW settings and a pure Chamfer Distance objective.PCN and GCN use CD-based regression; PDN and Mamba-DG use denoising score matching, while completion models use the same learning rate, weight decay, and batch size.

5 Results and Discussion

Denoising improves reconstruction, but the best denoiser depends on the data domain, while temporal conditioning primarily improves Chamfer Distance rather than strict point-level overlap. Trait accuracy also improves substantially after delta-conditioning and trait-aware fine-tuning, although maize and mesh-derived measurements retain important error sources.

  • Denoising ablation: 63.3% and 60.7% CD reductions are obtained for Adaptive Temporal PoinTr and Vanilla PoinTr, respectively, after denoising real scans.Adaptive Temporal PoinTr changes from 0.0180 to 0.0066, while Vanilla PoinTr changes from 0.0191 to 0.0075.
  • Denoising ablation: The optimal denoiser is domain-dependent: GCN achieves Test CD = 0.0066 on Pheno4D, while Mamba-DG achieves Test CD = 0.0061 on SynthCrop4D.The passage relates GCN’s advantage to structured laser-triangulation artifacts and Mamba-DG’s advantage to procedural Gaussian noise with σ=0.01.
  • Temporal completion: Temporal PoinTr achieves lower real-data CD than Vanilla PoinTr across denoiser pairings, with ranges of 0.0066–0.0072 versus 0.0075–0.0088.The improvement is reported as statistically significant with paired t-test p < 0.0001 across 15 bootstrap rounds.
  • Temporal completion: Real-data F-Scores remain near-tied, while synthetic F-Scores can favor Vanilla PoinTr because prior-stage geometry is offset from the current stage by growth.On real data, Vanilla PoinTr + Mamba-DG reaches 0.2080; on synthetic data, Vanilla + Mamba-DG reaches 0.1823 versus 0.1482 for Temporal + Mamba-DG.
  • Temporal completion: The temporal prior provides a plant-specific structural cue by using the same plant’s previous-stage scan as cross-attention context.This differs from prior approaches using learned query tokens or RGB images as auxiliary context.
  • Trait accuracy: Real trait error is linked to sensor noise and domain differences, with relative height error of 25.7% on real data versus 9.6% on synthetic data.Hull volume has the highest real relative error at 53.8%, partly because stale prior geometry inflates the convex hull.
  • Trait accuracy: 71.5%, 70.3%, and 81.9% reductions are reported for height, hull-volume, and surface-area MAE after the two trait-focused modifications.The corresponding real-data changes are 0.277→0.079 m, 0.343→0.102 m3, and 1.033→0.187 m2.
  • Species-wise traits: Tomato reaches 6.6% height and 5.1% canopy-width relative error, while maize height improves from 19.2% to 7.9%.Maize hull-volume and surface-area errors remain 22.7% and 24.4%, respectively, due to unresolved thin-leaf geometry issues.

6 Conclusion

The paper presents a denoising-aware temporal pipeline and SynthCrop4D benchmark for recovering crop architecture from noisy, partially occluded scans. Results support treating denoising and temporal completion as a coupled, domain-aware process, while several computational and measurement limitations remain.

  • Contributions: SynthCrop4D provides paired clean/noisy point clouds for 20 maize and tomato plants across six growth stages.The dataset is procedurally generated and supports controlled reconstruction experiments.
  • Findings: Denoising choice accounts for a 63.3% reduction in reconstruction error, with GCN optimal for real noise and Mamba-DG for synthetic noise.This supports domain-aware selection of the denoiser within the reconstruction pipeline.
  • Findings: V2 delta-conditioning and trait-aware fine-tuning reduce real-world trait MAE by 37–82% across four biological measurements.The reported traits are height, canopy width, hull volume, and surface area.
  • Limitations: Mamba-DG sampling scales as O(n^2) at large point counts, and single-step inference was not systematically benchmarked.These limitations bound scalability and the evidence for faster inference.
  • Limitations: Poisson reconstruction does not produce watertight meshes on the scans, leaving surface area and LAI as approximate proxies.The conclusion identifies these measurements as approximate rather than exact.
  • Reproducibility: The datasets and implementation resources are publicly available through open-source repositories, procedural generation scripts, and GitHub code.The paper states that SynthCrop4D can be reproduced using the provided scripts and that model weights are available.
Loading 2608.28343v1…