Uncertainty and Feature-Based Weighted Loss for 3D Wheat Part

Segmentation

R. Reena

1 a

, John H. Doonan

2 b

, Kevin Williams

, Fiona M. K. Corke

, Huaizhong Zhang

and

Yonghuai Liu

Edge Hill University, Lancashire, U.K.

National Plant Phenomics Centre, Aberystwyth University, U.K.

Keywords:

Plant Phenotyping, 3D Point Cloud, Wheat, Part Segmentation.

Abstract:

Deep learning techniques and point clouds have proved their efﬁcacy in 3D segmentation tasks of objects.

Nevertheless, the accurate plant organ segmentation is a formidable challenge due to their complex structure

and variability. Furthermore, presence of over-represented and under-represented parts, occlusion, and uneven

distribution complicates the 3D part segmentation tasks. Even though deep learning techniques often exhibit

exceptional performance, they also face challenges in applications where accurate trait estimation is required.

To handle these issues, we propose a novel uncertainty and feature based weighted loss that incorporates un-

certainty metrics and features of the plant or crop. We use Gradient Attention Module (GAM) with PointNet++

baseline to validate our approach. By dynamically introducing uncertainty and feature scores into the training

process, it promotes more balanced learning. Through comprehensive evaluation, we illustrate the advantages

of UFL (Uncertainty and Feature based Loss) as compared to standard CE (Cross entropy loss) with our own

constructed real Wheat dataset. The outcomes demonstrate consistent improvements in Accuracy (ranging

from 0.9% to 4.2%) and Ear mIoU (ranging from 1.8% to 15.3%) over the standard Cross-Entropy (CE) loss

function. As a result, our work contributes to the development of more robust and reliable segmentation mod-

els. This approach not only pushes forward the boundaries of precision agriculture but also has the potential

to inﬂuence related areas where accurate segmentation is pivotal.

1 INTRODUCTION

Component phenotyping refers to the measurements

of different plant organs i.e., leaf and stem length, leaf

width, leaf area etc. These measurements are essen-

tial markers for tracking plant development and fore-

casting yield (Shen et al., 2024). The extraction of

morphological and architectural traits are critical for

high throughput plant phenotyping, which requires

precise segmentation of plant organs. Manual meth-

ods are dominant, quite labor intensive, and unable

to meet the requirements of analysis of large pop-

ulations (Minervini et al., 2015). In the past few

years, computer vision and deep learning techniques

have facilitated plant phenotyping (Mochida et al.,

2018). With the development of point cloud based

deep learning networks i.e., PointNet (Qi et al., 2016)

and its enhanced variant PointNet++ (Qi et al., 2017),

https://orcid.org/0009-0007-9642-4607

https://orcid.org/0000-0001-6027-1919

remarkable progress has been made in 3D point-based

deep neural networks. Besides its proliferation, cur-

rent deep learning approaches on point cloud data

are usually limited to synthetic datasets i.e. Mod-

elNet40 and ShapeNet (Chang et al., 2015), S3DIS

(Armeni et al., 2016) and their application of 3D real

world data for plant segmentation tasks is at its in-

fancy. One of the underlying reasons for their lim-

ited application in plant domain is the lack of rich

annotated datasets and their labeling process takes a

lot of manual power (Chaudhury et al., 2020). Real

datasets, particularly those concerning plants, exhibit

signiﬁcant differences from their synthetic counter-

parts. In response, different architectures have been

developed for plant segmentation tasks (Turgut et al.,

2022), (Ghahremani et al., 2021), (Li et al., 2022).

These efforts highlight the ongoing advances in tai-

loring segmentation models to the unique challenges

presented by real agricultural data. Despite these in-

novations, achieving precise estimations remains a

challenge. Incorrect classiﬁcation leads to erroneous

632

Reena, R., Doonan, J. H., Williams, K., Corke, F. M. K., Zhang, H. and Liu, Y.

Uncertainty and Feature-Based Weighted Loss for 3D Wheat Part Segmentation.

DOI: 10.5220/0013312300003912

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages

632-641

ISBN: 978-989-758-728-3; ISSN: 2184-4321

estimates of the traits, which could signiﬁcantly inﬂu-

ence agricultural strategies and outcomes.

Previous research studies (Leibig et al., 2016;

Mukhoti et al., 2021; Kong et al., 2023) demonstrate

the signiﬁcance of uncertainty quantiﬁcation in im-

proving the robustness of model predictions. For in-

stance, (Wang et al., 2020) discusses incorporation of

uncertainty measure into their technique signiﬁcantly

enhances both the accuracy and effectiveness of their

approach. Similarly, uncertainty quantiﬁcation con-

cept is employed to enhance segmentation tasks in au-

tonomous driving applications, ensuring both safety

and reliability (Sun et al., 2024; Landgraf et al.,

2024). The uncertainty estimates are distinct from

softmax probabilities, by revealing the model’s con-

ﬁdence in its predictions. These estimates help iden-

tify potential uncertainties within the data that could

lead to incorrect outcomes. Though utilization of un-

certainty is explored in segmentation tasks, it is under

explored in the agricultural domain as compared to

its frequent application in medical and automotive do-

mains. In this work we propose a novel UFL (Uncer-

tainty and Feature based Loss) to bridge this gap that

utilizes uncertainty estimation during training and in-

tegrate features of the wheat crop into the loss func-

tion to enhance the segmentation of the models. Pri-

marily, we leverage the model’s uncertainty estimates

to quantify the conﬁdence levels of its predictions and

then incorporate important features. By integrating

feature importance analysis with uncertainty quantiﬁ-

cation into a hybrid scoring mechanism within our

loss function, our approach adeptly identiﬁes critical

features, such as wheat ear (spike) counts, that inﬂu-

ence model predictions, while simultaneously quanti-

fying the associated uncertainties. We conduct exten-

sive experiments with our constructed Wheat dataset

that has three different varieties i.e. Paragon, Gladius

and Apogee to exhibit the effectiveness of our tech-

nique using GAM (Hu et al., 2023). The motivation is

that plants have quite complex structures and they var-

ied according to their environmental conditions. Our

ﬁndings demonstrate that innovative integration of the

loss function and the critical features signiﬁcantly en-

hances the robustness of segmentation tasks. Precise

and accurate segmentation is quite crucial in plants

to make estimates about yield and making different

agricultural decisions. Lastly, we exemplify ablation

studies, limitations and future work of our study.

2 RELATED WORK

In the domain of segmentation and 3D object detec-

tion, numerous research studies have employed the

concept of uncertainty. We offer a summary of sig-

niﬁcant studies that utilize uncertainty aware tech-

niques for computer vision tasks across different do-

mains. Over past few decades uncertainty estimation

has been prevailing topic of researchers in image pro-

cessing tasks. (Blundell et al., 2015) introduced the

use of a Gaussian distribution over network weights,

characterizing each weight with a mean and variance.

However, the discussed technique was not optimal

for deep learning architectures due to a large number

of model parameters. With similar motivation, (Gal

and Ghahramani, 2016) suggested using Monte Carlo

(MC) dropout to apply dropout techniques for vari-

ational inference. During training, neurons are ran-

domly dropped, creating a Bernoulli distribution over

the network weights. While testing, the same image is

fed through the network multiple times with dropout,

allowing for estimation of uncertainty from these out-

puts. Quantiﬁcation of uncertainty is exercised in seg-

mentation and object detection tasks. This approach

is especially critical in autonomous driving scenarios,

where safety is a critical issue, as incorrect predictions

can result in severe consequences.

(Landgraf et al., 2024) has addressed this issue

with proposition of uncertainty aware segmentation

that will make estimates about the uncertainty dur-

ing training. On the basis of measures of uncertainty,

pixel-wise weight is attached to the regular cross en-

tropy loss to improve the segmentation performance.

With similar aim, panoptic segmentation is achieved

based on calibration of per pixel uncertainty estimates

(Sirohi et al., 2022). This study has covered seman-

tic and instance segmentation. In another study, un-

certainty estimation and segmentation module is pro-

posed to boost the segmentation performance in un-

certain regions by exploiting the uncertainty maps

(Bian et al., 2020). To handle imbalance issue in

segmentation tasks, (Bischke et al., 2018) has sug-

gested adaptive uncertainty weighted class loss to im-

prove the outcomes of neural network architectures.

A fuzzy uncertainty based method has been proposed

to expose out of distribution data samples for se-

mantic segmentation (Lin et al., 2023). The above

discussed technique integrates test-time augmentation

and Monte Carlo dropout with a trained segmenta-

tion model to generate multiple predictions, capturing

both data and model uncertainties. These predictions

form interval fuzzy sets, which are used to calculate

an image-level uncertainty score for identifying out-

of-distribution (OOD) samples. (Sun et al., 2024) has

explored label uncertainty or annotated uncertainty

information, for precise object detection in 3D data.

In the ﬁeld of healthcare, (Leibig et al., 2016) has

utilized uncertainty information for disease detection.

Uncertainty and Feature-Based Weighted Loss for 3D Wheat Part Segmentation

633

Paragon Gladius Apogee

Figure 1: Annotated 3D models of three wheat cultivars.

Authors have employed the concept of Bayesian un-

certainty that works better than other alternatives. The

ﬁndings indicate that incorporating uncertainty data

into decision-making processes can enhance diagnos-

tic accuracy. Another study has employed uncertainty

map to address the issue of unreliable segmentation

(Tang et al., 2022). To effectively extract meaningful

features and harness the potential of uncertainty, the

researchers introduced an uncertainty-guided module.

This module improves the model’s ability to discern

subtle features that are often overlooked in standard

segmentation approaches. In addition to autonomous

driving and medical imaging, research has also been

conducted in synthetic aperture radar (SAR) applica-

tions (Haas and Rabus, 2021). This study evaluates

uncertainty measurements to assess real-world quality

control performance. It utilizes softmax scores for un-

certainty estimation, which have been shown to offer

more practical insights compared to Bayesian meth-

ods.

3 METHODOLOGY

3.1 Dataset Description

The three wheat varieties—Paragon, Gladius, and

Apogee were grown in an ecologically controlled

Smarthouse at the National Plant Phenomics Cen-

tre (NPPC), Aberystwyth University, under experi-

ment reference W048. The plants were cultivated

on a conveyor system that was computer controlled,

and the soil water levels were changed to sustain

either a moderate drought or well-watered circum-

stances. Every plant was given a unique barcode that

was connected to all of the data it had gathered and

put on an RFID-tagged carriage. Over the course of

44 days, a 360-degree imaging strategy obtained 74

multi-view photos per plant every day. The 3D mod-

els were reconstructed using open source photogram-

metry software COLMAP (Sch

onberger and Frahm,

2016; Sch

onberger et al., 2016). After 3D reconstruc-

tion, the point cloud data was labelled using package

CloudCompare (Girardeau-Montaut, 2012). The seg-

ment function in the package was employed to divide

the plant into two parts and then the parts were anno-

tated using the Scalar function. Figure 1 depicts the

labelled point clouds. The red color is labelled with

’1’ that represents ear part (spikes) and the blue part

is labelled with ’0’ representing non-ear (leaves and

stems) part. We have downsampled each sample to

2048 points for the sake of computational feasibility.

Table 1: Dataset descriptions for different wheat categories.

Variety Total Train Val Test

Paragon 385 308 24 53

Gladius 371 296 25 50

Apogee 547 436 32 79

Wheat (All cultivars) 1303 1040 81 182

Despite the fact that they belong to the Wheat cat-

egory, the three cultivars show quite different charac-

teristics and architectures:

• Paragon: Mostly made up of leaves with a few tall

ears rising above them.

• Gladius: The species distinguished by having

more ears that are partly hidden beneath the fo-

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

634

Figure 2: A schematic overview of the UFL training process, which incorporates predictive uncertainties from a Monte Carlo

Dropout (MC-Dropout) and Features of plant to improve segmentation performance.

Figure 3: Distribution of Ear and Non-Ear parts in three

wheat cultivars.

liage.

• Apogee: A super dwarf with a high ear-to-leaf ra-

tio that is known for its quick cycling.

Table 1 illustrates the wheat dataset with all train-

ing, validation, and test samples used in this research

study. Figure 3 illustrates the proportion of ear points

and non-ear points of each wheat variety. Paragon ex-

hibits a notably smaller proportion of ear points com-

pared to those of Gladius and Apogee. Gladius has a

reasonable number of ear points, while Apogee shows

a signiﬁcantly larger proportion, highlighting the vari-

ability and imbalance in ear point distribution across

these varieties.

3.2 Proposed Approach

The core idea of our approach is to incorporate predic-

tive uncertainties and plant speciﬁc feature together

into the training process to enhance the accuracy of

part segmentation of crops or plants. The method in-

troduces an uncertainty and feature based loss func-

tion, which enhances the deep learning architecture’s

ability to prioritize learning from samples with high

uncertainty or signiﬁcant features. The proposed ap-

proach allows the model to focus on more challenging

data, improving overall learning efﬁciency and model

robustness. This approach introduces two primary en-

hancements into the training process:

1. Hybrid Score Calculation: During training,

Monte Carlo Dropout is exploited that measures

the uncertainty (u

) in prediction of the samples.

A hybrid score is calculated that integrates fea-

ture (e

) i.e. ear count (wheat spikes) in our case.

This score ensures that samples with both high

uncertainty and signiﬁcant feature attributes are

weighted appropriately.

2. Weighted Loss: We employ hybrid score as

weights in the loss function so that the model is

encouraged to focus on more uncertain and fea-

ture rich samples to facilitate effective learning.

Figure 2 represents the UFL computation process.

3.2.1 Architecture

We have adopted Gradient Attention module (GAM)

(Hu et al., 2023) with baseline PointNet++ for point

cloud analysis for our proposed technique. We inte-

grate the Gradient Attention Module (GAM), which

is designed to focus the network’s attention on cru-

cial features within the data, potentially improving

the accuracy and efﬁciency of segmentation. The

incorporation of GAM with PointNet++ aims to ex-

plore whether this attention mechanism can reﬁne the

Uncertainty and Feature-Based Weighted Loss for 3D Wheat Part Segmentation

635

model’s output by providing more focused and rel-

evant feature analysis, especially for complex plant

structures. The network utilizes gradient informa-

tion in the neighborhood and converts that informa-

tion into explicit representation. It is the ﬁrst to incor-

porate gradient data into the vector of locally aggre-

gated descriptors for point cloud neighborhood fea-

tures. The results show that the module has effec-

tively boost performance of the different state-of-the-

art methods.

3.2.2 Uncertainty and Feature Based Weighted

Loss

In contrast to the conventional utilization of Monte

Carlo Dropout, our proposed approach extends its ap-

plicability to the training phase. During each batch of

the training phase, we execute λ iterations to produce

λ segmentation samples in order to measure uncer-

tainties of the samples. The essential idea is that the

differences between the model’s successive predic-

tions reﬂect its conﬁdence level, which is a measure of

epistemic uncertainty. The high variation in these pre-

dictions indicates a reduced conﬁdence in the model’s

outputs, implying that the data poses challenges for

accurate identiﬁcation by the model. This variability

acts as a direct signal of uncertainty, making it pos-

sible to identify samples with signiﬁcant uncertainty.

Gradients are deactivated during this computational

phase to prevent them from affecting the backward

propagation process. The ultimate goal of UFL is to

enhance the standard categorical cross-entropy loss,

which is articulated as follows:

The modiﬁed uncertainty feature weighted loss

function can be expressed as:

UFL

= −

∑

i=1

∑

j=1

i, j

log(q

i, j

where:

• M is the total number of samples,

• K is the number of classes,

• w

represents the weight for the i-th sample,

• y

i, j

is the ground truth probability for the i-th sam-

ple and j-th class,

• q

i, j

is the predicted probability for the i-th sample

and j-th class.

Sample Wise Weight:

In our technique, the sample weight w

is inﬂuenced

by uncertainty scores and the plant features, notably

the ’ear’ count of wheat. The weight for each sample

is formulated as follows:

= α · e

+ β · u

where:

• α and β are parameters that modulate the contri-

butions of the ’ear’ feature importance (e

) and un-

certainty (u

) respectively.

• e

represents the ’ear’ feature importance for sam-

ple i

• u

denotes the uncertainty associated with sample

i, reﬂecting the conﬁdence level of the measure-

ments or predictions.

and u

have quite disparity in the scale so we have

chosen min-max normalization to ensure that these in-

puts are comparable and can contribute appropriately

to the model’s decision making process. This step will

enhance the network’s ability to integrate these met-

rics effectively, leading to more accurate and reliable

segmentation predictions. α and β control the weights

of the loss function. We have presented experiments

to investigate the impact of these parameters on the

performance of deep learning neural network archi-

tecture in the next section.

3.3 Experiments

In this section, we undertake a comprehensive series

of experiments to validate the effectiveness of inte-

grating predictive uncertainties into the training pro-

cess.

3.3.1 Experimental Set up

Training: The network is trained for 150 epochs

with a batch size of 8 with 2048 points in each point

cloud. All the experiments are conducted utilizing

the PyTorch deep learning framework, leveraging the

computational power of an NVIDIA GPU with 12

GB VRAM. For all training procedures, we utilize a

Stochastic Gradient Descent (SGD) (Robbins, 1951)

optimizer conﬁgured with an initial base learning rate

of 0.001, a momentum of 0.9, and a weight decay

of 0.0001. The number of segmentation sample λ is

set to 5 by default. All experiments are evaluated on

Paragon, Gladius and Apogee i.e. Wheat varieties.

For quantitative evaluations, we primarily report the

mean Intersection over Union (mIoU), and Accuracy.

3.3.2 Quantitative Results

We have employed GAM module with PointNet++

(Qi et al., 2017) as baseline.

Table 2 shows a notable improvement in segmen-

tation accuracy and mean Intersection over Union

(mIoU) when we apply our Uncertainty and feature

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

636

Table 2: Comparative Study of Segmentation Methods Across Three Datasets.

Dataset GAM (Baseline PointNet++) Accuracy mIoU Non-Ear mIoU Ear mIoU

Paragon

CE 0.917 0.650 0.907 0.392

UFL α = 1.25, β = 0.75 0.959 0.750 0.956 0.545

UFL α = 1.5, β = 1 0.951 0.685 0.942 0.422

Gladius

CE 0.892 0.712 0.870 0.554

UFL α = 1.25, β = 0.75 0.919 0.771 0.904 0.637

UFL α = 1.5, β = 1 0.898 0.741 0.880 0.603

Apogee

CE 0.908 0.818 0.869 0.768

UFL α = 1.25, β = 0.75 0.917 0.831 0.884 0.782

UFL α = 1.5, β = 1 0.915 0.830 0.879 0.783

(a) Paragon Dataset (b) Gladius Dataset

Figure 4: Model Accuracy Comparison: GAM with baseline PointNet++ vs. Our Method for three different datasets (Paragon,

Gladius, Apogee). Each graph shows the segmentation accuracy across the training epochs.

based loss (UFL) as compared to traditional Cross En-

tropy (CE) loss across all the datasets. Particularly α

= 1.25 and β = 0.75 consistently outperforms other

conﬁgurations, suggesting that optimal balance be-

tween uncertainty and feature values assist in model

learning. This conﬁguration leads to substantial gains

in both Ear mIoU and Non-Ear mIoU, underlining

its effectiveness in differentiating between more and

less challenging segmentation areas. Paragon accu-

racy rose from 0.917 to 0.959 and mIoU witnessed a

jump of 10% in the ﬁgures. Gladius saw slight im-

provements in Accuracy but there is rise of 6% in

mIoU. Although Apogee is the one wheat variety that

is quite complex and the results reﬁned from 0.908

accuracy to 0.917 but the Ear mIoU has a slight de-

cline as compared to the scenario when α = 1.5 and

β = 1, highlighting the trade-offs involved in optimiz-

ing these parameters. Figure 4 presents the training

Uncertainty and Feature-Based Weighted Loss for 3D Wheat Part Segmentation

637

(a) Paragon (Base Method) (b) Paragon (Ours Method , α = 1.25, β = 0.75)

(e) Apogee (Base Method) (f) Apogee (Ours Method, α = 1.25, β = 0.75)

Figure 5: Comparison of segmentation results on the Paragon, Gladius, and Apogee datasets. Each dataset shows a baseline

and our method view to illustrate the effectiveness of the proposed strategy. In the visual representations, the ear part is

marked in red, while the non-ear part is depicted in green for both the Ground Truth and Prediction columns. Correct

predictions are highlighted in cyan, whereas magenta is used to indicate misclassiﬁcations. This color-coded approach helps

to clearly delineate the improvements our method brings in accurately segmenting the ear and non-ear components across

different wheat varieties.

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

638

Table 3: Segmentation Methods Across Three Datasets with GTNet.

Dataset GTNet Accuracy mIoU Non-Ear mIoU Ear mIoU

Paragon

CE 0.927 0.662 0.923 0.400

UFL α = 1.5, β = 1 0.932 0.676 0.929 0.425

UFL α = 1.25, β = 0.75 0.931 0.658 0.929 0.388

Gladius

CE 0.886 0.667 0.873 0.460

UFL α = 1.5, β = 1 0.881 0.721 0.854 0.588

UFL α = 1.25, β = 0.75 0.862 0.660 0.839 0.481

Apogee

CE 0.811 0.668 0.751 0.584

UFL α = 1.5, β = 1 0.817 0.688 0.742 0.633

UFL α = 1.25, β = 0.75 0.819 0.686 0.748 0.628

accuracy of our method with each variety. The ﬁg-

ures demonstrate that our proposed approach yields

a marked improvement in accuracy over the baseline

method. This enhancement is evident across vari-

ous metrics and scenarios, highlighting the effective-

ness of integrating feature importance and uncertainty

quantiﬁcation into the segmentation process.

Furthermore, an increase in values of α =1.5 and

β = 1 leads to decline in certain performance met-

rics as compared to previous discussed parameter set-

tings but the overall results consistently surpass those

achieved using the traditional Cross-Entropy (CE)

loss. This trend is particularly evident across multiple

datasets, where even with higher parameter values,

the UFL method demonstrates superior capability in

handling complex segmentation tasks. This suggests

that careful tuning of these parameters can lead to op-

timized model performance enabling more accurate

segmentation.

3.3.3 Qualitative Results

Figure 5 presents qualitative results of baseline and

our proposed approach. We have chosen the best re-

sults of our technique. It is clear from the results for

Paragon and Gladius that the number of misclassiﬁ-

cations has decreased. Although it is more challeng-

ing to assess the impact for Apogee due to its com-

plex wheat variety, quantitative data indicate substan-

tial improvements for Apogee as well.

3.3.4 Comparative Study with GTNet

To further demonstrate its potential, we carry out a

comparative study between our proposed technique

with our deep learning network architecture with GT-

Net (Zhou et al., 2024), which is another network ar-

chitecture that integrates the Transformer and Graphs

based approaches. It features a Local Transformer,

which utilizes intra-domain cross-attention, and a

Global Transformer that employs global self-attention

to extract local and global features efﬁciently. We

have opted this model due to its strength in segmen-

tation tasks. Table 3 illustrates the impact of our loss

function as compared to the basic CE loss function.

Performance in Paragon has improved with α =1.5

and β = 1 with the average mIoU experiencing a slight

increase from 0.662 to 0.667 but the Ear mIoU saw

a slight dip. Similarly, Gladius has shown enhanced

performance under the same settings, with its mIoU

rising from 0.667 to 0.721. Apogee also displays a

similar trend in performance, with its mIoU improv-

ing from 0.668 to 0.688 as compared to the baseline

loss function. However the segmentation accuracy in

each variety has a slight improvement. For instance,

Paragon accuracy have a rise from 0.927 to 0.932 but

this is not the same with Gladius which has seen a

certain drop in this metric. Apogee is one variety that

has seen enhancement in the results of segmentation

accuracy from 0.811 to 0.817. We have observed that

while both parameter settings are effective, they do

not achieve results as impressive as those obtained

with GAM under similar conditions.

3.4 Ablation Studies

In addition to the quantitative and qualitative results

above, we also perform multiple ablation studies be-

low.

Impact of α and β : The α controls uncertainty

scores and β have an impact on feature i.e. ear count

in our case. In this section we have evaluated the im-

pact of α and β with varied values to explore their im-

pact on the model’s performance. These parameters

are crucial to understand the sensitivity of our deep

learning architecture to these adjustments. Interest-

ingly, some parameter settings may lead to decrease in

performance. Table 4 presents the performance of the

architecture with our proposed technique. α =1 and β

=1 consistently outperformed the CE baseline across

all datasets. But with another setting ( α =1.4 and β

Uncertainty and Feature-Based Weighted Loss for 3D Wheat Part Segmentation

639

Table 4: Ablation studies of Segmentation Methods Across Three Datasets.

Dataset GAM (Baseline PointNet++ Accuracy mIoU Non-Ear mIoU Ear mIoU

Paragon

CE 0.917 0.650 0.907 0.392

UFL α = 1, β = 1 0.954 0.714 0.951 0.477

UFL α = 1.4, β = 0.9 0.949 0.699 0.946 0.453

Gladius

CE 0.892 0.712 0.870 0.554

UFL α = 1, β = 1 0.911 0.764 0.895 0.634

UFL α = 1.4, β = 0.9 0.888 0.706 0.873 0.539

Apogee

CE 0.908 0.818 0.869 0.768

UFL α = 1, β = 1 0.910 0.830 0.883 0.776

UFL α = 1.4, β = 0.9 0.910 0.818 0.876 0.761

= 0.9) of optimized parameters the performances face

a slight decrease especially in case of Apogee. These

ﬁndings highlight a critical balance in the parameter

settings within our UFL framework.

Increasing each parameter beyond threshold can

impact model’s performance. So an optimal parame-

ter setting is pivotal that provides substantial improve-

ments over baseline without compromising model’s

robustness. In this ablation a small change in param-

eters have an inﬂuence on segmentation accuracy and

mIoU. To exemplify this, when α has an increase of

0.4 and β has very small change of 0.1 the perfor-

mance of the model decreases in each wheat variety.

Paragon saw a decline of 5% in segmentation accu-

racy and a similar trend is followed in Gladius where

accuracy dropped from 0.892 to 0.888. Though all

varieties have surpassed the baseline values but there

is small fall in the Ear mIoU of Apogee from 0.768 to

0.761 in comparison to the baseline CE loss.

4 DISCUSSION

In this research study, we have observed noteworthy

enhancements in segmentation accuracy and mIoU

following speciﬁc adjustments to parameters α and

β across different wheat cultivars datasets. These

improvements strengthen our hypothesis that uncer-

tainty and feature based loss with ﬁne tuning of α

and β can enhance model performance, illustrating

a complex interplay between model robustness and

data speciﬁcity. Particularly, the difference in mag-

nitude of performance across datasets underscores

the inﬂuence of dataset-speciﬁc characteristics on

model efﬁcacy, challenging the viability of a uniform

parameter-tuning approach across different scenarios.

Limitations: The suggested approach might not be

as effective for different crops that exhibit varying

structures and densities. Furthermore, there is a

risk that elevated values of α and β could lead to

overﬁtting, compromising the model’s generalization

across diverse agricultural scenarios. The general-

izabiltity of this approach needs to be yet explored

with different plant and crop species. Moreover, we

rely only on ear count as sole feature in this study.

We can consider different features i.e. ear ratio,

ear height, ear weight so they can also contribute to

better learning.

5 CONCLUSIONS

In conclusion, our study has demonstrated that un-

certainty and features scores are contributing with the

ﬁne tuning of parameters α and β in our UFL frame-

work that show notable performance enhancement of

segmentation tasks with different wheat varieties. We

have discerned marked improvements in segmenta-

tion accuracy and mIoU, which underscores the po-

tential of our proposed technique. The results exhibit

that a uniform parameter setting might be less effec-

tive across diverse crop species and the parameters

tuning needs to be adjusted on speciﬁc data character-

istics can lead to substantial improvements. This not

only highlights the adaptability of our model but also

points to the importance of dataset-speciﬁc strategies

in agricultural applications. With the precise consid-

eration of α and β the risk of overﬁtting can also be

mitigated. This paves the way for more robust, ac-

curate, and efﬁcient models that can be pivotal in ad-

vancing precision agriculture.

REFERENCES

Armeni, I., Sener, O., Zamir, A. R., Jiang, H., Brilakis, I.,

Fischer, M., and Savarese, S. (2016). 3d semantic

parsing of large-scale indoor spaces. In 2016 IEEE

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

640

Conference on Computer Vision and Pattern Recogni-

tion (CVPR), pages 1534–1543.

Bian, C., Yuan, C., Wang, J., Li, M., Yang, X., and and,

S. Y. (2020). Uncertainty-aware domain alignment

for anatomical structure segmentation. Medical Im-

age Analysis, 64:101732.

Bischke, B., Helber, P., Borth, D., and Dengel, A. (2018).

Segmentation of imbalanced classes in satellite im-

agery using adaptive uncertainty weighted class loss.

In IGARSS 2018 - 2018 IEEE International Geo-

science and Remote Sensing Symposium, pages 6191–

6194.

Blundell, C., Cornebise, J., Kavukcuoglu, K., and Wierstra,

D. (2015). Weight uncertainty in neural networks.

Chang, A. X., Funkhouser, T., Guibas, L., Hanrahan, P.,

Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S.,

Su, H., Xiao, J., Yi, L., and Yu, F. (2015). Shapenet:

An information-rich 3d model repository.

Chaudhury, A., Boudon, F., and Godin, C. (2020). 3d plant

phenotyping: All you need is labelled point cloud

data. In Workshop on Computer Vision Problems in

Plant Phenotyping, pages 244–260.

Gal, Y. and Ghahramani, Z. (2016). Dropout as a bayesian

approximation: Representing model uncertainty in

deep learning.

Ghahremani, M., Williams, K., Corke, F. M. K., Tiddeman,

B., Liu, Y., and Doonan, J. H. (2021). Deep segmen-

tation of point clouds of wheat. Frontiers in Plant

Science, 12.

Girardeau-Montaut, D. (2012). CloudCompare - Open

Source project — cloudcompare.org. https://www.

cloudcompare.org/. [Accessed 17-11-2023].

Haas, J. and Rabus, B. (2021). Uncertainty estimation

for deep learning-based segmentation of roads in syn-

thetic aperture radar imagery. Remote Sensing, 13(8).

Hu, H., Wang, F., Su, J., Zhou, H., Wang, Y., Hu, L., Zhang,

Y., and Zhang, Z. (2023). Gam : Gradient attention

module of optimization for point clouds analysis. In

AAAI Conference on Artiﬁcial Intelligence.

Kong, L., Kamarthi, H., Chen, P., Prakash, B. A., and

Zhang, C. (2023). Uncertainty quantiﬁcation in deep

learning. In Proceedings of the 29th ACM SIGKDD

Conference on Knowledge Discovery and Data Min-

ing, KDD ’23, page 5809–5810, New York, NY, USA.

Association for Computing Machinery.

Landgraf, S., Wursthorn, K., Hillemann, M., and Ulrich,

M. (2024). Dudes: Deep uncertainty distillation using

ensembles for semantic segmentation. PFG – Journal

of Photogrammetry, Remote Sensing and Geoinforma-

tion Science, 92(2):101–114.

Leibig, C., Allken, V., Ayhan, M. S., Berens, P., and Wahl,

S. (2016). Leveraging uncertainty information from

deep neural networks for disease detection. Scientiﬁc

Reports, 7.

Li, D., Shi, G., Li, J., Chen, Y., Zhang, S., Xiang, S., and

Jin, S. (2022). Plantnet: A dual-function point cloud

segmentation network for multiple plant species. IS-

PRS Journal of Photogrammetry and Remote Sensing,

184:243–263.

Lin, Q., Chen, X., Chen, C., Pekaslan, D., and Garibaldi,

J. M. (2023). Fuzzy uncertainty-based out-of-

distribution detection algorithm for semantic segmen-

tation. In 2023 IEEE International Conference on

Fuzzy Systems (FUZZ), pages 1–6.

Minervini, M., Scharr, H., and Tsaftaris, S. (2015). Im-

age analysis: The new bottleneck in plant phenotyping

[applications corner]. IEEE Signal Processing Maga-

zine, 32:126–131.

Mochida, K., Koda, S., Inoue, K., Hirayama, T., Tanaka, S.,

Nishii, R., and Melgani, F. (2018). Computer vision-

based phenotyping for improvement of plant produc-

tivity: a machine learning perspective. GigaScience,

8(1):giy153.

Mukhoti, J., Kirsch, A., van Amersfoort, J. R., Torr, P. H. S.,

and Gal, Y. (2021). Deep deterministic uncertainty:

A new simple baseline. 2023 IEEE/CVF Conference

on Computer Vision and Pattern Recognition (CVPR),

pages 24384–24394.

Qi, C., Yi, L., Su, H., and Guibas, L. J. (2017). Pointnet++:

Deep hierarchical feature learning on point sets in a

metric space. In NIPS, page 5105–5114.

Qi, C. R., Su, H., Mo, K., and Guibas, L. J. (2016). Pointnet:

Deep learning on point sets for 3d classiﬁcation and

segmentation. arXiv preprint arXiv:1612.00593.

Robbins, H. E. (1951). A stochastic approximation method.

Annals of Mathematical Statistics, 22:400–407.

Sch

onberger, J. and Frahm, J.-M. (2016). Structure-from-

motion revisited.

Sch

onberger, J., Zheng, E., Pollefeys, M., and Frahm, J.-

M. (2016). Pixelwise view selection for unstructured

multi-view stereo. volume 9907.

Shen, J., Wu, T., Zhao, J., Wu, Z., Huang, Y., Gao, P., and

Zhang, L. (2024). Organ segmentation and phenotypic

trait extraction of cotton seedling point clouds based

on a 3d lightweight network. Agronomy, 14(5).

Sirohi, K., Marvi, S., B

uscher, D., and Burgard, W. (2022).

Uncertainty-aware panoptic segmentation.

Sun, Y., Lu, B., Liu, Y., Yang, Z., Behera, A., Song, R.,

Yuan, H., and Jiang, H. (2024). Exploiting label un-

certainty for enhanced 3d object detection from point

clouds. IEEE Transactions on Intelligent Transporta-

tion Systems, 25(6):6074–6089.

Tang, P., Yang, P., Nie, D., Wu, X., Zhou, J., and Wang,

Y. (2022). Uniﬁed medical image segmentation by

learning from uncertainty in an end-to-end manner.

Knowledge-Based Systems, 241:108215.

Turgut, K., Dutagaci, H., and Rousseau, D. (2022). Ros-

esegnet: An attention-based deep learning architec-

ture for organ segmentation of plants. Biosystems En-

gineering, 221:138–153.

Wang, J., Ancha, S., Chen, Y.-T., and Held, D. (2020).

Uncertainty-aware self-supervised 3d data associa-

tion. In 2020 IEEE/RSJ International Conference on

Intelligent Robots and Systems (IROS), pages 8125–

8132.

Zhou, W., Wang, Q., Jin, W., Shi, X., and He, Y. (2024).

Graph transformer for 3d point clouds classiﬁcation

and semantic segmentation. Computers & Graphics,

124:104050.

Uncertainty and Feature-Based Weighted Loss for 3D Wheat Part Segmentation

641