Plant organ segmentation on compact computing platforms requires accurate predictions and affordable geometric preparation. We develop and evaluate a multiscale edge based graph attention model that combines local shape descriptors, two neighborhood scales and a residual stem and leaf correction in approximately 78 thousand parameters. The engineering application is plant point cloud segmentation for precision agriculture, assessed through grouped accuracy experiments and physical Raspberry Pi 5 inference. Broad comparisons with hierarchical, graph and transformer baselines establish PointNeXt as the strongest external accuracy reference among the tested local recipes. A nested four variant ablation on a local maize collection separates geometric information from the combined multiscale configuration across seven outer groups and three seeds. The complete model achieves the highest mean intersection over union, 86.36%, although uncertainty and boundary errors qualify the contribution of each component. We then optimize full cloud principal component analysis, sampling and neighbor queries without changing learned weights or descriptor definitions. All 1,180 paired desktop comparisons on PlantNet and TomatoWUR preserve exact logits and predictions. Multiscale pipeline latency falls from 330.33 to 96.82 milliseconds on PlantNet and from 1,054.52 to 463.10 milliseconds on TomatoWUR. PointNeXt also benefits from optimized sampling and remains faster over the complete pipeline. These results support compact graph learning with reduced geometric overhead, while physical board measurements establish inference feasibility rather than energy efficiency or readiness for autonomous flight.
Figures & tables
Aspect
Earlier preprint ( Ruiz-Martinez et al., 2025 )
Present study
Shared core
XYZ, ten PCA channels, coordinate kNN, EdgeConv and GAT
Same broad representation and established operator families
Configuration
Neighborhood size 16
K8/K16 backbone; separate organ logit correction
Evaluation
Fold validation on maize
Historical collection comparisons; nested AO factorial with explicit provenance limits
Implementation
No paired P4 or ARM study reported
Frozen prediction checks and stage timings on desktop and Pi; three repeated ARM cycles on PlantNet/TomatoWUR
Table 1: Scope relative to the earlier preprint version, arXiv:2507.00182v2. Differences describe reported methods and evidence, not a matched performance comparison.
Figure 1: Multiscale EdgeGAT. (a) Full-cloud PCA descriptors, sampling and coordinate neighborhoods supply two multiscale edge blocks and two graph attention layers. P4 optimizes preparation without changing the trained network. A separate branch combines local refinement and global context to correct only stem and leaf logits; argmax assigns the final labels. (b) Each edge block computes relative inputs separately over nested neighborhoods of 8 and 16 points. The geometric term rij(k) uses a scale specific local normalization. Independently learned messages are aggregated by neighbor maxima and fused into 32 features per point. Blue marks the nearest eight neighbors; ochre marks the next eight, also included at k=16 . The plant and neighborhoods show real geometry, not predicted labels. LN denotes layer normalization. The editable vector schematic was prepared with OpenAI Codex assistance.
Collection
Clouds
Groups
Classes
Group basis
AO
304
7
Ternary
Filename prefixes
Pheno4D subset
120
40
Ternary
Local clusters
ZeaMays
2,080
520
Ternary
Source identifiers
Single Maize
428
5
Binary
Cultivar holdouts
PLANesT–3D
34
31
Binary
Reconstruction groups
PlantNet
546
35
Binary
Species and replicates
Table 2: Local collection structure. Counts refer to files and protocol groups, not necessarily independent plants. Ternary labels are stem, soil and leaf; binary labels are stem and leaf.
Model
AO
Pheno4D
ZeaMays
PointNet++
86.90
75.88
84.94
DGCNN
87.06
78.15
89.60
Point Transformer
87.58
75.49
88.93
RandLA-Net
83.77
74.27
84.30
PSegNet XYZ
88.10
76.71
88.20
PSegNet + PCA
87.04
70.99
87.56
Table 3: Broad ternary benchmark: mIoU in percent at seed 42, using all seven AO folds and all five folds on Pheno4D and ZeaMays. Checkpoints are validation selected. External recipes appear above the rule and our model configurations below it. Bold marks the highest mean in each column. PSegNet with PCA is a feature control.
Model
Single Maize
PLANesT 3D
PointNet++
88.35
69.85
DGCNN
94.48
77.08
Point Transformer
90.03
72.67
RandLA-Net
86.27
66.82
PSegNet XYZ
93.61
76.78
PointNeXt
95.34
83.01
Table 4: Broad binary benchmark: mIoU in percent at seed 42, with all five folds per model and collection. These are validation selected checkpoints, distinct from the nested AO evaluation. Bold marks the highest mean in each column.
Collection
Model
mIoU
Stem IoU
Leaf IoU
PlantNet
PointNeXt
90.35
85.07
95.64
PlantNet
PTv3 adapter
88.28
81.89
94.67
PlantNet
PointCT adapter
63.01
43.61
82.41
PlantNet
Multiscale
90.39
85.15
95.63
TomatoWUR
PointNeXt
80.53
67.88
93.18
TomatoWUR
PTv3 adapter
77.00
61.85
92.15
Table 5: Plant extension: all four evaluated recipes, each with five folds and three seeds (42, 52 and 62), giving 15 fold results per row. Scores are percentages from validation selected checkpoints. Bold marks the highest mIoU per collection. The adapters use local training and compatibility settings.
Figure 2: Overview of the complete available baseline cohorts. Panel A uses seed 42 with seven AO folds and five folds elsewhere; its first three collections are ternary and its last two binary. Panel B shows the separate binary plant extension with five folds and three seeds. Cells contain mIoU in percent; N/E means not evaluated in that cohort. Colors use one fixed scale, without pooling datasets, tasks or seed cohorts. Generated from archived CSV tables with an AI assisted plotting script.
Collection
Classes
Runs
Multiscale
PointNeXt
Δ (pp)
AO
3
21
88.29
88.71
-0.41
Pheno4D
3
15
75.84
80.00
-4.17
ZeaMays
3
15
88.67
90.47
-1.80
PLANesT-3D
2
15
80.63
82.69
-2.06
Single Maize
2
15
94.84
95.37
-0.53
PlantNet
2
15
90.39
90.35
+0.03
Table 6: Focused three seed comparison after the broad benchmark. Historical validation selected mIoU in percent, using all folds and seeds 42, 52 and 62. Runs are results per model. Differences use unrounded means. The PlantNet and TomatoWUR rows reuse the preceding extension cohort, rather than adding independent evidence.
Model
mIoU
Stem IoU
Soil IoU
Leaf IoU
E1
86.42
66.64
98.97
93.66
Multiscale
86.74
66.84
99.36
94.04
DGCNN
86.48
66.39
99.39
93.66
RandLA-Net
80.90
53.98
97.90
90.80
PointNeXt
86.73
67.22
99.40
93.56
Table 7: Broader nested AO comparison at seed 42 only. Seven outer groups contribute to each row; inner groups select checkpoints. Overlapping rows in the following three seed table are not additional independent evidence.
Model
mIoU
Stem IoU
Soil IoU
Leaf IoU
E1
85.71
65.18
98.75
93.20
Multiscale
86.36
66.55
98.84
93.68
PointNeXt
86.76
67.06
99.54
93.69
Table 8: Nested AO accuracy in percent, averaging seven outer groups and three paired seeds. Checkpoints are selected on separate inner groups.
Comparison
Measure
Difference (pp)
95% interval
E1 − Multiscale
mIoU
-0.64
[-1.68, 0.39]
E1 − Multiscale
Stem IoU
-1.36
[-3.97, 1.21]
E1 − Multiscale
Boundary error
0.51
[-1.14, 2.41]
Multiscale − PointNeXt
mIoU
-0.41
[-2.65, 2.46]
Multiscale − PointNeXt
Stem IoU
-0.51
[-5.55, 5.44]
Multiscale − PointNeXt
Boundary error
4.08
[1.46, 7.12]
Table 9: Descriptive paired AO differences in percentage points, with 95% group bootstrap intervals. Seeds are averaged before resampling. Positive boundary error differences are unfavorable.
Model
Configuration
Params.
mIoU
Stem IoU
Leaf IoU
Boundary error
E1
Single scale, XYZ
50,535
85.71
65.18
93.20
27.91
E2
Multiscale, XYZ
77,863
85.64
65.59
92.91
26.10
E3
Single scale, PCA
50,535
85.62
64.89
93.17
29.17
Multiscale
Multiscale, PCA
77,863
86.36
66.55
93.68
27.40
Table 10: Completed nested AO factorial ablation. Values are percentages averaged over seven outer groups and three seeds. Lower boundary error is better. E1 and Multiscale reuse the controls reported above; the rows are not additional independent evidence. XYZ variants retain ten constant zero descriptor channels during training.
Dataset
Model
mIoU ± SD
Stem IoU
Leaf IoU
PlantNet
Multiscale
90.34 ± 2.61
85.10
95.59
PlantNet
PointNeXt
90.30 ± 2.31
84.96
95.65
TomatoWUR
Multiscale
74.33 ± 1.80
57.55
91.11
TomatoWUR
PointNeXt
80.50 ± 1.59
67.87
93.13
Table 11: Seed 42 CPU accuracy under original and optimized preparation. Both arms have identical values. Means and standard deviations use five folds; organ columns are mean IoU.
Dataset
Model
Original ms
Optimized ms
Speedup
PlantNet
Multiscale
330.33
96.82
3.41 ×
PlantNet
PointNeXt
277.93
82.34
3.38 ×
TomatoWUR
Multiscale
1,054.52
463.10
2.28 ×
TomatoWUR
PointNeXt
334.89
123.26
2.72 ×
Table 12: Paired desktop pipeline latency, in milliseconds per cloud. PointNeXt optimization affects external sampling only; it has no external PCA or graph preparation in this protocol. One cycle is measured on four physical CPU cores.
Figure 3: Measured Multiscale desktop cost by stage, using separate scales for the two datasets. Other processing includes reading, normalization, feature assembly, packing, label selection and timed overhead. PCA retains the complete raw cloud support. Generated from recorded timings with an AI assisted plotting script.
Figure 4: Complete desktop pipeline means for both models, with original and optimized preparation. PointNeXt receives the compatible sampling optimization. Its absence of PCA is part of its actual recipe, not a missing timing component. Generated from recorded timings with an AI assisted plotting script.
Dataset
Model
Network ms
Prep. ms
Sum ms
RSS MiB
PlantNet
Multiscale
88.70
673.97
765.47
383.55
PlantNet
PointNeXt
262.72
469.74
732.86
386.59
TomatoWUR
Multiscale
90.97
2,498.75
2,590.63
426.13
TomatoWUR
PointNeXt
260.75
611.83
873.14
378.88
Table 13: Completed Raspberry Pi 5 reference campaign without P4. Network is the mean of five fold profile medians; preparation and preparation plus network are means over clouds. The sum omits packing and postprocessing and must not be directly subtracted from the new desktop total. RSS includes the audit runtime.
Collection
Model
Control ms
P4 ms
Speedup
Reduction (%)
P4 p50 ms
P4 p95 ms
PlantNet
Multiscale
760.00
277.73
2.74 ×
63.46
254.45
438.05
PlantNet
PointNeXt
722.51
342.47
2.11 ×
52.60
345.31
372.03
TomatoWUR
Multiscale
2,612.96
1,196.28
2.18 ×
54.22
1,179.17
1,592.76
TomatoWUR
PointNeXt
877.72
458.51
1.91 ×
47.76
457.18
503.76
Table 14: Paired Raspberry Pi 5 pipeline latency. Control and P4 are means across clouds of per-cloud medians over three cycles, in ms. Speedup is the ratio of those means. P4 p50 and p95 are percentiles across the per-cloud medians. Each row uses five folds and training seed 42.
Figure 5: Measured original and P4 pipeline latency on Raspberry Pi 5, with separate scales for each collection. Bars are means across clouds of per-cloud medians over three repeated cycles, not distributions over independent plants or training runs. PointNeXt receives optimized sampling. Generated from the verified ARM timing CSV with an AI assisted plotting script.
Collection
Model
Preparation
Mean of cycle means
SD of cycle means
Range of cycle means
PlantNet
Multiscale
Original
761.19
4.88
755.65–764.87
PlantNet
Multiscale
P4
278.94
0.32
278.68–279.29
PlantNet
PointNeXt
Original
723.48
1.51
721.96–724.98
PlantNet
PointNeXt
P4
342.65
1.99
340.36–343.97
TomatoWUR
Multiscale
Original
2,602.77
16.63
2,583.95–2,615.47
TomatoWUR
Multiscale
P4
1,195.89
4.37
1,190.98–1,199.37
Table 15: Descriptive variation across three ARM cycle means, in ms. For each cycle, latency is first averaged across clouds; the table then summarizes those three means. This aggregation differs from Table 14 , so its SD must not be attached to that table’s principal mean-of-medians. Cycles repeat the same inference workload.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Collection
Multiscale
LCCA
LCCA − Multiscale
GLR1 − Multiscale
AO
88.29
88.78
+0.48
-0.07
Pheno4D
75.84
75.67
-0.16
+0.09
ZeaMays
88.67
88.71
+0.04
-0.06
Appendix
Table A.1: Historical architectural alternatives. LCCA denotes the late cloud context adapter; GLR1 is the geometric logit residual. GLR1 differences use its own paired controls and cannot be added to another campaign’s absolute means.
3D point cloud segmentation assigns semantic, instance, or part labels to points for scene understanding, mapping, robotics, and scientific measurement.
Recent 3D foundation models provide powerful feature representations for point cloud learning by controlling spatial granularity. However, relying on a fixed spatial granularity severely limits generalization in applications like plant phenotyping, where organ morphology and size vary substantially across species and growth stages. To address this, we propose AGS-PlantSeg, a few-shot 3D plant organ segmentation method that leverages the frozen Utonia (arXiv:2603.03283) foundation model combined with Adaptive Granularity Selection. By dynamically selecting the best granularity levels for each specific plant model, our method extracts optimized geometric features for a lightweight MLP segmentation head. Extensive experiments across PLANesT-3D (arXiv:2407.21150), Pheno4D , and Crops3D demonstrate that AGS-PlantSeg significantly improves cross-species generalization, achieving 88.9% average mIoU performance and outperforming fixed-granularity baselines by 2.5 mIoU points. Despite requiring minimal annotated data, our approach is highly competitive with fully supervised, plant-specific architectures.
Carla Salazar, Lazaros Nalpantidis
Technical University of Denmark (DTU), Kongens Lyngby, Denmark · Pioneer Centre for Artificial Intelligence, Copenhagen, Denmark
Accurate 3D plant organ segmentation is fundamental to automated phenotyping. Existing approaches rely on annotated training data or species-specific model configurations. We present an annotation-free pipeline for 3D plant organ segmentation, combining text-prompted SAM3 segmentation with semantic neural radiance fields (NeRFs). Given only multi-view RGB images and a list of class names, our zero-shot pipeline produces semantically labeled 3D point clouds without manual annotation, per-species fine-tuning, or domain-specific preprocessing. Multi-view NeRF fusion acts as effective implicit consensus mechanism that lifts imperfect per-frame masks into accurate 3D labels. On a controlled Begonia maculata testbed the SAM3 pipeline achieves 92.6% mIoU, reaching 95.9% of the oracle upper bound established with perfect ground-truth masks. The pipeline was further evaluated on a new dataset spanning ten diverse plant point clouds reaching an average 0.856 mIoU, with leaf and pot IoU above 0.91 and 0.90 for every species, respectively. These results demonstrate that annotation-free 3D plant organ segmentation is now feasible and approaching the range of supervised methods.
Andreas Gilson, Laura Hennig, Peter Pietrzyk
Fraunhofer Institute for Integrated Circuits (IIS), Fürth, Germany · Otto-Friedrich-Universität Bamberg, Germany · Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany
The accuracy of existing leaf-wood segmentation methods for tree point clouds varies across forest types and sites. Self-supervised learning (SSL) on point clouds has improved the generalization of deep learning models for forestry point cloud tasks, including biomass regression and individual tree segmentation, but its applicability to leaf-wood segmentation remains untested. In this study, we pretrained Point-M2AE, a widely used SSL architecture for point clouds, on ShapeNet-55 augmented with 2,400 individual tree point clouds. For fine-tuning and inference, we used recursive voxel subdivision to handle the wide variation in point density across inputs, allowing the same model to operate at both individual-tree and plot scales without architecture change. Compared to the model without pretraining, the pretrained model improved wood IoU from 60.5% to 70.0% for needleleaf and from 69.7% to 76.3% for broadleaf trees. On a benchmark spanning four countries across three climatic zones, the pretrained model achieved the smallest cross-site variation and highest overall performance among compared methods (LeWos, CWLS, and PointTransformer). Plot-level segmentation maintained accuracy comparable to individual-tree performance, with mIoU of 84.7% for broadleaf and 77.7% for needleleaf plots, showing that the model generalizes across scales without additional finetuning. As a downstream test in tropical forests, where dense canopies make segmentation challenging, we applied our model and a quantitative structure model to estimate wood volume for 28 trees from Guyana, Indonesia, and Peru to assess whether the segmentation improvements from SSL pretraining translate into improved downstream performance. The resulting volume estimates achieved the lowest error among all methods tested (MAE = 2.40 m3), less than half that of algorithmic baselines (LeWos: 5.94 m3; CWLS: 5.27 m3).
Heeju Mun, Tackang Yang, Yunsoo Nam +1
Department of Landscape Architecture and Rural Systems Engineering, Seoul National University, Seoul, South Korea · Interdisciplinary Program in Landscape Architecture, Seoul National University, South Korea · Integrated Major in Smart City Global Convergence, Seoul National University, Seoul, Republic of Korea +1