DCSFormer: a high-precision method for cotton seedling point cloud organ segmentation

TL;DR

This paper introduces DCSFormer, a new method for accurately segmenting cotton seedling organs in 3D point clouds, improving precision and generalization.

Contribution

DCSFormer introduces a novel DCS Block and CLFSkip for better organ segmentation in cotton seedlings using point clouds.

Findings

DCSFormer achieves 93.67% mIoU, 95.83% mPrec, 97.35% mRec, and 96.56% mF1 in cotton seedling segmentation.

The model outperforms baseline models on Crops3D and Pheno4D datasets across all metrics.

A new annotated cotton seedling dataset was created to support training and evaluation.

Abstract

Accurately segmenting cotton seedling organs from 3D point clouds is fundamental for high-throughput plant phenotyping and digital breeding. However, cotton seedling segmentation remains challenging due to fine-scale and complex organ morphology, uneven point density with noise, and the lack of high-quality annotated datasets.

To address these issues, we propose DCSFormer, a tailored extension of Point Transformer V3 designed for cotton seedling point cloud segmentation. The model introduces the DCS Block, which leverages dynamic sparse expert routing and dual-channel attention to adaptively capture global semantic dependencies and subtle local geometric variations, thereby improving stem-leaf boundary discrimination. In addition, the proposed CLFSkip replaces traditional skip connections with a cross-layer fusion strategy, effectively integrating multi-scale features while preserving organ-level details. We also constructed an annotated cotton seedling dataset to support training and evaluation.

Experimental results show that DCSFormer achieves 93.67% mIoU, 95.83% mPrec, 97.35% mRec, and 96.56% mF1, outperforming multiple comparison models. Furthermore, when evaluated against baseline models on two public datasets, Crops3D and Pheno4D, DCSFormer exceeds the baseline across all four metrics, further validating its effectiveness and generalizability. This work provides an effective solution for precise cotton seedling organ segmentation.

Figures