3D Instance Segmentation · Volumes of Irregular Shapes · Phenotyping

Computer Vision powered Wheat Phenotyping

We developed a novel 3D instance segmentation model to analyze wheat plant structures from point cloud data. Our method achieves state-of-the-art accuracy while being efficient enough for real-world deployment in agricultural settings. We create a novel volume estimation dataset for accurate wheat head volume estimation.

Dataset WheatFormer3D Datasets
Framework PyTorch · MMDetection3D
Hardware A100 GPU
Task 3D Segmentation · Volume Regression · Object Size Estimation

Problem Statement

Efficient agricultural practices can be greatly benefited by accurate phenotyping of crops, which involves measuring traits like size and volume of plant structures. Traditional manual phenotyping is labor-intensive and error-prone, while 2D image-based methods struggle with occlusions and complex geometries. We aim to develop a 3D instance segmentation model that can accurately identify and measure wheat heads from point cloud data, enabling high-throughput phenotyping for improved crop management and breeding.

Method Overview

We develop a transformer-decoder based architecture, WheatFormer3D, and domain specific augmentations for robust 3D instance segmentation. Our model is designed to be efficient and robust under annotation sparsity, making it suitable for real-world agricultural applications. We also create a novel volume estimation dataset, WheatHeadVol3D, and propose a regression-based approach for accurate volume estimation of irregularly shaped wheat heads.

Visual Results

3D Instance Segmentation Result

3D Semantic Segmentation Result

Wheat head size measurement

Wheat head size measurement

Wheat head size measurement

WheatFormer3D and WheatHeadVol3D pipelines.

Wheat scanning setup

3D scanning setup for wheat head reference volumes.

Quantitative Results - 3D Instance Segmentation

Average Precision (AP) metrics computed at various IoU thresholds. For the baselines, we report metrics that are evaluated with their official implementation. Prec50 and Rec50 are missing for Soft Group.

Method AP25 AP50 AP Prec50 Recall50
WheatFormer3D 91.19 87.96 77.99 96.14 85.38
Oneformer3D 80.52 76.48 65.63 96.55 73.93
Oneformer3D-P 85.99 78.21 63.24 93.87 72.87
Mask3D 82.31 76.06 63.41 92.89 83.99
SoftGroup 40.84 33.08 26.86 - -

Quantitative Results - Volume Regression

Volume estimation metrics from WheatHeadVol3D test set for both pooling strategies, max-pooling and transformer-decoder. In-field: Samples from in-field scans of wheat plots. Indoor: Corresponding high-res indoor scans of wheat heads. MAE in cm3.

Type MAE MAPE ρ
In-fieldIndoor In-fieldIndoor In-fieldIndoor
Max-pooling 1.350.22 18.194.26 0.720.99
Transformer-decoder 0.870.19 11.563.71 0.810.99
Convex-hull 26.3910.20 380.16187.75 0.450.97

Quantitative Results - Trait Analysis

Measurement results for length (L), width (W), volume (convex-hull) (V) and volume (regression model) (V2). MAE in cm for L and W and cm3 for V and V2. p-value ≪ 0.001 for all measurements.

Score
Threshold
IoU #
Measured
MAE MAPE ρ
LWVV2 LWVV2 LWVV2
0.30.3291 0.980.131.770.29 9.619.9924.1714.69 0.850.540.730.78
0.50.5273 0.890.111.570.26 8.669.0022.5912.80 0.860.570.750.82
0.70.7234 0.800.081.230.24 7.485.8213.6611.20 0.880.620.870.84

Key Contributions