Scale Matters - AGS-PlantSeg

Abstract

We introduce AGS-PlantSeg, a few-shot 3D plant organ segmentation method built on a frozen Utonia encoder with a lightweight two-layer MLP head. Rather than committing to a single fixed granularity, our Adaptive Granularity Selection (AGS) module adapts the representation to each plant. During training, AGS selects three granularity levels using prototype-based measures of inter-class separation, intra-class compactness, and class-wise boundary consistency: one level maximizes class discriminability, while two class-specific levels capture boundary consistency. At inference, it selects three levels based on similarity to stored training prototypes and combines their predictions through confidence-weighted aggregation. Trained only on PLANesT-3D and evaluated for cross-species, cross-dataset generalization on Pheno4D and Crops3D, AGS-PlantSeg reaches 88.9% average mIoU, outperforming the best fixed-granularity baseline (86.4%) by 2.5 mIoU.

88.9%

avg. mIoU, AGS-PlantSeg (few-shot)

86.4%

best fixed-granularity baseline

+2.5

mIoU improvement

Overview

A single fixed spatial granularity rarely suits every plant. AGS-PlantSeg adapts granularity per plant instead, trained only on PLANesT-3D and evaluated for cross-species, cross-dataset generalization on Pheno4D and Crops3D.

Effect of granularity/scale selection on leaf and stem point segmentation for Ribes and Rose plants — Effect of granularity on per-point leaf/stem segmentation. Ribes 07 prefers the finer granularity (Scale=7); Rose 06 prefers the coarser granularity (Scale=1): no single fixed scale suits both.

Frozen Utonia features

Point features come from a frozen Utonia encoder at multiple granularities; only a lightweight two-layer MLP head is trained.

→

Adaptive selection (AGS)

Per plant, pick 3 granularity levels: training uses prototype scores (inter-class separation, intra-class compactness, boundary consistency); inference uses similarity to stored prototypes.

→

Confidence-weighted aggregation

Combine the 3 levels' predictions into the final per-point organ label (e.g. stem, leaf).

Method

AGS selects granularity levels using label prototypes, both when training the segmentation head and at inference on unseen plants.

Contributions

01
Adaptive Granularity Selection (AGS). Per plant, not per region: at training time AGS selects three granularity levels (one discriminability-optimal level, plus one boundary-aware level per class) via prototype scores of inter-class separation, intra-class compactness, and class-wise boundary consistency; at inference it selects three levels by prototype similarity and combines them via confidence-weighted aggregation.
02
Few-shot, frozen-feature setting. AGS-PlantSeg builds on a frozen Utonia encoder with only a lightweight two-layer MLP head trained, no backbone fine-tuning.
03
Cross-species, cross-dataset generalization. Trained only on PLANesT-3D and evaluated on unseen Pheno4D and Crops3D (tomato, potato), AGS-PlantSeg reaches 88.9% avg. mIoU vs. 86.4% for the best fixed-granularity baseline.

Quantitative Results

mIoU vs. fixed granularity per species, compared against AGS-PlantSeg's adaptive selection — Per-species mIoU across fixed granularities (1–7) vs. AGS-PlantSeg (adaptive, ours). No single fixed scale wins across species; adaptive selection matches or beats the best fixed scale for every species.

Distribution of AGS-selected granularity scale per species, shown as violin plots — Distribution of the scale AGS actually selects per species, across all test plants — confirming the module adapts genuinely per species rather than converging to one scale.

BibTeX

If you find this work useful, please cite: