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.
Figures & tables
Figure 1: The necessity of adaptive granularity in plant phenotyping. Cosine similarity ( green : low, red : high) for leaves (top) and stems (bottom) between prototypes and Utonia features across different input scales for two example plants. A single fixed granularity fails to generalize across varying plant morphologies. The first plant requires a much finer granularity (scale=7) compared to the second (scale=1).
Figure 2: Overview of AGS-PlantSeg. The method consists of five stages: (a) prototype computation at a candidate granularity level g , (b) training-time granularity selection using prototype-based scores, (c) segmentation head training on the selected granularity levels Gtrain∗ , (d) inference-time granularity selection using stored training prototypes and predicted labels, and (e) final prediction from the selected granularity levels Ginfer∗ . Orange and blue denote the two semantic organ labels. Feature spaces are visualized with PCA for illustration.
Method
Supervision
Pepper
Rose
Ribes
Average
Rel. Training Speed
Acc
mIoU
Acc
mIoU
Acc
mIoU
Acc
mIoU
PointNet++ [ 7 ]
Full
97.9
94.3
96.7
89.8
98.4
94.2
97.7
92.8
-
RoseSegNet [ 7 ]
Full
98.3
95.5
97.4
92.0
98.6
95.1
98.1
94.2
-
SP-LSCnet [ 7 ]
Full
98.1
95.0
97.1
89.8
98.3
94.5
97.8
93.1
-
GCASSN [ 24 ]
Full
95.6
97.1
92.8
95.4
93.7
96.5
94.0
96.3
-
Ours
Full
98.8
96.9
98.5
94.4
99.2
97.5
98.8
96.3
1.0 ×
Table 1: Comparison with baseline methods on PLANesT-3D (%) . PointNet++, RoseSegNet, and SP-LSCnet are reported from PLANesT-3D [ 7 ] , using the same 70/30 plant-level split as our full setting. GCASSN [ 24 ] uses a different 80/20 split and preprocessing protocol. Best result is shown in bold and second-best is underlined.
Method
Supervision
Pheno4D
Crops3D
Avg
Tomato
Tomato
Potato
Acc
mIoU
Acc
mIoU
Acc
mIoU
Acc
mIoU
GCASSN [ 24 ]
Full
91.5
60.8
-
-
-
-
-
-
Ours
Full
97.1
86.0
94.3
82.6
97.2
85.6
96.2
84.7
Ours
Few-Shot
96.6
83.7
93.6
81.3
96.7
83.6
95.6
82.9
Ours
One-Shot
94.0
81.4
93.2
80.2
97.5
85.9
94.9
82.5
Table 2: Generalization to unseen plant species and datasets (%) . AGS-PlantSeg is evaluated with full, few-shot, and one-shot supervision. Baseline results are reported as published in the corresponding papers. Best result is shown in bold and second-best is underlined.
Figure 3: Qualitative generalization results on unseen species and datasets. Predictions are shown alongside ground truth for Pheno4D tomato and Crops3D tomato and potato. Orange denotes leaves and blue denotes stems.
Scale
PLANesT-3D
Pheno4D
Crops3D
Avg
Pepper
Rose
Ribes
Tomato
Tomato
Potato
Acc
mIoU
Acc
mIoU
Acc
mIoU
Acc
mIoU
Acc
mIoU
Acc
mIoU
Acc
mIoU
1.0
97.4
92.7
96.1
86.7
97.3
87.7
93.8
70.6
92.0
77.2
97.1
84.8
95.6
83.3
2.0
98.3
95.2
97.5
91.0
98.3
92.3
89.5
74.8
93.8
81.2
95.7
79.0
95.5
85.6
3.0
98.7
96.4
97.4
90.7
98.8
94.5
95.1
83.4
91.1
76.4
94.3
77.2
95.9
86.4
4.0
98.8
96.8
97.6
91.5
98.9
94.9
95.6
83.0
91.1
76.5
92.0
71.3
95.7
85.7
Table 3: Comparison with fixed-granularity alternatives in the few-shot setting. Fixed baselines use a single Utonia scale for both training and inference, while AGS selects granularity levels adaptively. Best result is shown in bold and second-best is underlined.
Method
BFP
Sboundaryg,c
CWA
Avg. mIoU
Ablation 1
87.3
Ablation 2
✓
86.8
Ablation 3
✓
88.3
Ablation 4
✓
✓
88.5
Ours
✓
✓
✓
88.9
Table 4: Ablation of AGS components in the few-shot setting. BFP denotes boundary-filtered prototype computation. Sboundaryg,c indicates whether boundary-aware granularity scores are used. CWA denotes confidence-weighted aggregation at inference time. Best result is shown in bold and second-best is underlined.
Northwest A&F University, Yangling, 712100, China · Huazhong University of Science and Technology, Wuhan, 430074, China · Wuhan Institute of Technology, Wuhan, 430205, China