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

Mrudul Mittal; Soumyashree Kar

cs.CVAug 28, 2026

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

Authors: Mrudul Mittal, Soumyashree Kar

Organizations: Department of Electronics Engineering, Sardar Vallabhbhai National Institute of Technology, Surat, Gujarat, India · Centre of Studies in Resources Engineering, Indian Institute of Technology Bombay, Mumbai, Maharashtra, India

Abstract

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.

Figures & tables

Figure 1: Overview of the dual-case evaluation pipeline. Real (Pheno4D) and synthetic (SynthCrop4D) point clouds are preprocessed into a unified normalized representation, then evaluated under two regimes: Case 1 , where completion architectures operate directly on raw, undenoised input, and Case 2 , where one of four denoising models (PCN, GCN, PDN, Mamba-DG) is applied before completion. Both cases use the $t{-}1$ growth stage as a temporal prior.

Figure 2: Temporal progression of real 3D plant point clouds from the Pheno4D dataset across different growth stages.

Figure 3: Temporal progression of procedurally generated 3D plant point clouds from the SynthCrop4D dataset across different growth stages.

Property	Pheno4D (Real)	SynthCrop4D (Synthetic)
Species	Maize, tomato	Maize, tomato (matched)
Temporal structure	Genuine $t{-}1\!\to\!t$ growth sequences	Genuine $t{-}1\!\to\!t$ growth sequences
Growth stages	5–6	6
Points per scan (after preprocessing)	4,000	4,000
Split protocol	By plant identity (no temporal leakage)	By plant index (no temporal leakage)
Perturbation magnitude	Registration residual, $\sim\!\sigma{=}0.01$ scale	Gaussian, $\sigma{=}0.01$ (matched)

Table 1: Correspondence and divergence between Pheno4D and SynthCrop4D. Rows above the midrule indicate deliberately matched properties, enabling direct cross-domain comparison; rows below indicate properties that necessarily differ by construction and motivate SynthCrop4D’s role as a controlled complement to, rather than a replacement for, real data.

Figure 4: Architecture of the proposed Mamba-DG denoiser. The model operates in three stages: (1) 3D points are serialized via Hilbert Encoding and processed through Mamba SSM blocks to capture global context; (2) SSM computation uses ZOH discretization and a pure PyTorch parallel prefix scan; (3) features are scattered back to 3D space, combined with local $k$ -NN frames, and fed into a ScoreNet followed by Langevin dynamics for iterative denoising.

Model	Mechanism	Loss	Notes
PCN	PointNet $\to$ displacement	CD-L1	Local only
GCN	EdgeConv on $k$ -NN graph	CD-L1	Feature-space graph
PDN	PointNet + UOF $\to$ Langevin	DSM	Pretrained
Mamba-DG	Hilbert SSM $\to$ Langevin	DSM	$\mathcal{O}(n)$ global

Table 2: Denoising architectures compared (Section 3.3 ).

Model	Mechanism	Uses $t{-}1$	Notes
PCN	FC decoder on global desc.	No	Displaces points
Vanilla PoinTr	FPS proxy + self-attn	No	Ablation
Adaptive PoinTr	Self-attn + cross-attn	Yes	Ours

Table 3: Completion architectures compared (Section 3.4 ).

Figure 5: Architecture of the Adaptive Temporal PoinTr completion model. Both the current partial scan at $t$ and the complete prior scan at $t{-}1$ are encoded and downsampled into proxy tokens via FPS. A Spatial Transformer refines the current frame’s features; a Temporal Cross-Attention Decoder queries the prior scan’s geometry. A coarse-to-fine MLP decoder then generates the missing geometry.

Hyperparameter	PCN	GCN	PDN	Mamba-DG (Ours)
Loss	CD	CD	DSM	DSM
Optimiser	AdamW	AdamW	AdamW	AdamW
Learning rate	$1\times 10^{-3}$	$1\times 10^{-3}$	$5\times 10^{-5}$	$1\times 10^{-3}$
Weight decay	0	$1\times 10^{-4}$	0	0
LR schedule	Fixed (none)	Cosine annealing	Cosine annealing	Cosine annealing
Batch size	1	1	1	1

Table 4: Denoising-model training hyperparameters. Selected independently per architecture rather than via a shared grid search, since the four denoisers differ in loss paradigm, initialisation, and memory footprint.

Hyperparameter	PCN Baseline / Vanilla PoinTr / Adaptive Temporal PoinTr
Loss	Chamfer Distance (mean NN pred $\to$ GT + mean NN GT $\to$ pred); no auxiliary term
Optimiser	AdamW
Learning rate	$1\times 10^{-4}$ (single shared rate; no backbone/decoder split)
Weight decay	$1\times 10^{-4}$
Batch size	2
Epochs (max)	100

Table 5: Completion-model training hyperparameters. Unlike the denoisers, all three completion architectures share an identical recipe.

Model	Paradigm	Train CD $\downarrow$	Val CD $\downarrow$	Test CD $\downarrow$	F@5mm $\uparrow$
PCN	Displacement Reg. (CD)	0.0922	0.0899	0.0862	0.0453
GCN	Graph Conv. (CD)	0.0919	0.0896	0.0861	0.0472
PDN	Local DG (DSM)	0.0922	0.0899	0.0863	0.0471
Mamba-DG	Global SSM + DG	0.0919	0.0897	0.0860	0.0471

Table 6: Denoising results on 74 held-out Pheno4D test clouds (30-epoch checkpoints, 15,000 max points). While PCN and GCN train using CD and PDN/Mamba-DG use DSM, we report unified Chamfer Distances (CD) across all splits to ensure direct cross-comparability.

Completion Model	Input Source	Real Pheno4D			Synthetic SynthCrop4D
Completion Model	Input Source	Train	Val	Test	Train	Val	Test
PCN Baseline	Case 1 (noisy, no denoising)	0.0879		0.0800	0.0281		0.0384
	Case 2: Mamba-DG (Ours)	0.0263	0.0300	0.0286	0.0145	0.0148	0.0175
	Case 2: GCN	0.0288	0.0322	0.0306	0.0185	0.0165	0.0233
	Case 2: PDN	0.0271	0.0311	0.0298	0.0166	0.0154	0.0192
	Case 2: PCN	0.0277	0.0297	0.0299	0.0165	0.0165	0.0222

Table 7: Master ablation: Case 1 (no denoising) vs. Case 2 (with denoising) across 4 denoisers

\times

3 completion architectures on Pheno4D and SynthCrop4D. Metric: CD-L2

\downarrow

. Bold = best per test column.

Completion Model	Input Source	Real Pheno4D			Synthetic SynthCrop4D
Completion Model	Input Source	Train	Val	Test	Train	Val	Test
PCN Baseline	Case 1 (noisy, no denoising)	0.0051		0.0072	0.0124		0.0207
	Case 2: Mamba-DG (Ours)	0.0071	0.0069	0.0090	0.0123	0.0137	0.0119
	Case 2: GCN	0.0065	0.0056	0.0084	0.0106	0.0122	0.0119
	Case 2: PDN	0.0069	0.0061	0.0094	0.0117	0.0104	0.0116
	Case 2: PCN	0.0069	0.0065	0.0074	0.0110	0.0117	0.0119

Table 8: Master ablation: F-score@5 mm

\uparrow

for all configurations. Same experimental setup as Table 7 . Bold = best per test column. PCN Baseline F-scores near zero reflect the inability of a displacement-only decoder to hallucinate new geometry outside the partial input support.

Figure 6: Qualitative evaluation of 3D reconstruction performance on unseen Pheno4D test plants (Maize M06/M07 and Tomato T06/T07). Rows (top to bottom): (1) Raw noisy partial input; (2) PCN baseline reconstruction; (3) PoinTr baseline; (4) Ours (Adaptive Temporal PoinTr with denoising); (5) Ground truth (green). Denoising-augmented completion preserves delicate leaf topologies and avoids the structural collapse observed without denoising.

Method	Year	Venue	Dataset	Occlusion	CD $\downarrow$	F $\uparrow$	Temp.	Denoise	Trait
				Type			Prior	Stage	Valid.
Chen et al. [ 34 ]	2023	Plant Phen.	Custom RGB-D (cabbage)	Single-view depth	—	—	No	No	Area only
PCN [ 4 ]	2018	3DV	PlantCom3D	Depth-map proj.	0.804	0.512	No	No	No
PoinTr [ 5 ]	2021	ICCV	PlantCom3D	Depth-map proj.	0.367	0.850	No	No	No
SeedFormer [ 35 ]	2022	ECCV	PlantCom3D	Depth-map proj.	0.051	0.977	No	No	No
AdaPoinTr [ 36 ]	2023	T-PAMI	PlantCom3D	Depth-map proj.	0.077	0.953	No	No	No

Table 9: Contextual comparison with prior plant point cloud completion methods. All prior methods use synthetic datasets or RGB-D sensors; our method is evaluated on real laser-triangulation scan data (Pheno4D [ 7 ] ). PlantCom3D entries report CD-L1

\times 10^{3}

; our Pheno4D entries report raw CD-L2. Direct numerical comparison is not valid across datasets or metrics.

Panel of Figure 7 — Figure 7: Phenotypic trait extraction from ground-truth (green) and reconstructed (red) 3D point clouds. Top: raw point clouds. Middle: plant height and canopy width from the axis-aligned bounding box. Bottom: hull volume from a 3D convex hull.

Trait	Pheno4D		SynthCrop4D
Trait	MAE	%Err	MAE	%Err
Plant Height (m)	0.277	25.7%	0.060	9.6%
Canopy Width (m)	0.185	9.1%	0.096	22.1%
Hull Volume (m 3 )	0.343	53.8%	0.021	104.4%
Surface Area (m 2 )	1.033	26.9%	0.229	46.2%

Table 10: Baseline trait MAE on real Pheno4D (60 observations) and SynthCrop4D (15 observations). Relative error

=

MAE / mean GT.

Trait	Maize		Tomato
	Pred mean $\pm$ MAE	%Err	Pred mean $\pm$ MAE	%Err
Height (m)	$1.647\pm 0.280$	19.2%	$0.890\pm 0.275$	32.1%
Canopy Width (m)	$1.416\pm 0.210$	13.6%	$2.438\pm 0.171$	7.4%
Hull Volume (m 3 )	$0.623\pm 0.242$	61.5%	$1.101\pm 0.401$	51.5%
Surface Area (m 2 )	$3.962\pm 1.525$	49.5%	$4.138\pm 0.748$	17.5%
Eff. LAI	$1.044\pm 0.338$	46.9%	$0.694\pm 0.083$	11.9%

Table 11: Baseline species-wise trait accuracy on real Pheno4D. Tomato

n=38

, Maize

n=22

stage observations.

Trait	Baseline	V2 MAE	Improv.	V2 %Err	V2 $r$
Plant Height (m)	0.277	0.079	71.5%	7.3%	0.94
Canopy Width (m)	0.185	0.115	37.8%	5.7%	0.97
Hull Volume (m 3 )	0.343	0.102	70.3%	16.1%	0.84
Surface Area (m 2 )	1.033	0.187	81.9%	13.0%	0.90

Table 12: V2 vs. baseline trait MAE on real Pheno4D (60 observations). Pearson

r

: per-stage GT vs. predicted trajectory correlation.

Trait	Maize		Tomato
	Pred mean $\pm$ MAE	%Err	Pred mean $\pm$ MAE	%Err
Height (m)	$1.502\pm 0.116$	7.9%	$0.895\pm 0.057$	6.6%
Canopy Width (m)	$1.480\pm 0.113$	7.3%	$2.221\pm 0.117$	5.1%
Hull Volume (m 3 )	$0.396\pm 0.089$	22.7%	$0.725\pm 0.110$	14.2%
Surface Area (m 2 )	$1.002\pm 0.208$	24.4%	$1.684\pm 0.175$	9.8%

Table 13: V2 species-wise trait accuracy on real Pheno4D. Tomato

n=38

, Maize

n=22

Figure 8: V2 trait trajectories on real Pheno4D test plants. GT (solid) vs. V2 reconstructed (dashed red); shaded region $=\pm 1$ std. Each panel reports per-species MAE and Pearson $r$ . Rows (top to bottom): height, canopy width, hull volume, surface area. Left: Maize; right: Tomato.

Species	Trait	GT $R^{2}$	Pred $R^{2}$	GT $t_{0}$	Pred $t_{0}$	$\Delta t_{0}$
Maize	Height	0.32	0.35	0.9	1.8	$+0.9$
Tomato	Height	0.80	0.75	1.5	3.5	$+2.0$
Tomato	Hull Volume	0.41	0.29	4.1	3.5	$-0.6$
Tomato	Surface Area	0.22	0.19	2.1	3.4	$+1.3$

Table 14: V2 logistic growth curve

R^{2}

and change point

t_{0}

\Delta t_{0}=t_{0}^{\text{pred}}-t_{0}^{\text{GT}}

(positive

=

phase lag). Entries marked with a dash indicate

R^{2}<0.20

for both GT and predicted, where sigmoid fitting is not meaningful.

Figure 9: V2 logistic growth curve fits. GT (solid green) vs. predicted (dashed red). Change points $t_{0}$ marked with $\blacktriangledown$ . Rows: Maize and Tomato height (top); Tomato hull volume and surface area (bottom).

Point Cloud Denoising

See all 10 papers

Jul 2, 2026

The Turning Point of 3D Plant Phenotyping: 3D Foundation Models Enable Minute-to-Second Cross-Crop Reconstruction and Beyond

3D plant phenotyping is notoriously known to be procedure-complicated and of low throughput due to the extensive multi-view imaging, the fragile 3D reconstruction pipeline, and the additional cost from reconstructed geometry to phenotypic extraction. These limitations are further amplified in low-cost data acquisition, where smartphone videos or sparsely sampled multi-view images provide limited view overlap and self-occlusion. In this work, we show that the conventional 3D plant phenotyping pipeline could be streamlined and significantly accelerated with 3D Foundation Models (3DFMs), and particularly, present one of the first cross-crop 3D phenotyping frameworks powered by 3DFMs. The framework replaces COLMAP-style sparse initialization with 3DFM-based feed-forward geometric recovery, combines geometry-constrained 3D Gaussian Splatting for dense reconstruction, enables few-view reconstruction through iterative view synthesis and refinement, and converts reconstructed geometry into measurable organs through 2D-to-3D semantic transfer, metric scale recovery, and organ instance separation. We further construct a cross-crop dataset with smartphone-based image acquisition, diverse plant morphologies, and manual annotations for segmentation and phenotypic evaluation. Experiments across 26 plant sequences show that 3D Foundation Models reduce the average reconstruction time from 6.52 minutes to 1.58 seconds while maintaining high reconstruction quality and phenotyping accuracy. These results suggest a fresh technical route for high-throughput 3D plant phenotyping, from low-cost image acquisition to fast reconstruction, perception, scale recovery, and phenotypic measurement.

Hanyue Jia, Wei Zhou, Wenbo Zhou +3

Northwest A&F University, Yangling, 712100, China · Huazhong University of Science and Technology, Wuhan, 430074, China · Wuhan Institute of Technology, Wuhan, 430205, China

3D Foundation Models Plant Phenotyping Sparse-View 3D Reconstruction Feed-Forward 3D Reconstruction

Ask PDF

Sep 15, 2026

Evaluating Mesh Reconstruction Methods for Crop Phenotyping

Phenotyping an agricultural crop is crucial for studying its entire life cycle, as it provides vital insights to improve yield and, ultimately, food production. Doing the same for crops grown on remote sites is a challenge for the specialists who cannot be available on-site. 3D reconstruction techniques offer a promising solution to this problem by enabling crop digitization, allowing specialists to access the resulting 3D crop models from anywhere at any time. In this work, we evaluate recent 3D reconstruction pipelines for crop phenotyping. We focus on 7 mesh reconstruction pipelines and measure the fidelity and consistency of their outputs qualitatively and quantitatively. Our results suggest that the meshes produced by the GGGS, PGSR, and 2DGS are preferable to the other pipelines, owing to their quantitative metrics and visually pleasing outputs. The GGGS pipeline is better than the second-best pipeline (2DGS) by about 27% on the radar chart with 5 dimensions, namely, User ratings, Chamfer distance, LPIPS, PSNR, and SSIM.

Karanvir Singh, Theo Morales, Binh-Son Hua +1

Indian Institute of Technology Ropar, Rupnagar 140001, Punjab, India · Trinity College Dublin, The University of Dublin, College Green D02 PN40, Dublin 2, Ireland

3D Gaussian Splatting Plant Phenotyping 3D Reconstruction 3D Mesh Reconstruction

Ask PDF

Aug 8, 2026

SUMI: Scalable Unified Model for 3D Point Cloud Inference

Point cloud completion commonly follows a coarse-to-fine paradigm, where a low-density coarse shape is first predicted and then upsampled to the target resolution. Although recent methods have improved global structure recovery, the fine stage often remains limited by simple upsampling and insufficient interaction with coarse structural features, making local detail reconstruction challenging. We propose SUMI, a diffusion-enhanced refinement module for coarse-to-fine point cloud completion. Unlike prior diffusion-based completion methods that use diffusion as a standalone point generator, SUMI injects noisy geometric features into cross-attention with coarse structural features, enabling reverse denoising to refine local geometry while preserving global consistency. SUMI can also be integrated into existing coarse-to-fine models as a flexible refinement module. Experiments on PCN, ShapeNet-55/34, and MVP demonstrate consistent improvements over strong baselines. SUMI achieves the best overall CD and F1-score on PCN, reduces CD by up to 16.1% on ShapeNet-55, and obtains the best CD across all output densities on MVP.

Yanlong LI, Kanchana Thilakarathna

Point Cloud Completion 3D Shape Completion 3D Diffusion Models

Ask PDF

Abstract

Point Cloud Denoising

Explore similar work

The Turning Point of 3D Plant Phenotyping: 3D Foundation Models Enable Minute-to-Second Cross-Crop Reconstruction and Beyond

Evaluating Mesh Reconstruction Methods for Crop Phenotyping

SUMI: Scalable Unified Model for 3D Point Cloud Inference