A Robust Comparative Study of Adaptative Reprojection Fusion

Methods for Deep Learning Based Detection Tasks

with RGB-Thermal Images

Enrique Heredia-Aguado

, Marcos Alfaro-P

erez

, Mar

ıa Flores

, Luis Paya

David Valiente

and Arturo Gil

University Institute for Engineering Research, Miguel Hern

andez University, Avda. de la Universidad s/n,

03202 Elche (Alicante), Spain

Keywords:

Image Fusion, Multispectral Image, Object Detection, RGB-Thermal, Deep Learning.

Abstract:

Fusing visible and thermal imagery is a promising approach for robust object detection in challenging envi-

ronments, taking advantage of the strengths from different spectral information. Building on previous work

in static early fusion, we present a comparative study of adaptative reprojection fusion methods that exploit

advanced projections and frequency-domain transforms to combine RGB and thermal data. We evaluate Prin-

cipal Component Analysis, Factor Analysis, Wavelet and Curvelet-based fusion, all integrated into a YOLOv8

detection pipeline. Experiments are conducted on the LLVIP dataset, with a focus on methodological rigour

and reproducibility. This research show promising results based on these methods comparing to previous early

fusion methods. We discuss the implications for future research and the value of robust experimental design

for advancing the state of the art in multispectral fusion.

1 INTRODUCTION

Multimodal data fusion has consistently demonstrated

its value across various knowledge domains. By com-

bining different sources of information, it is possi-

ble to compensate for the limitations of individual

modalities and take advantage their complementary

strengths.

The primary application scenario for this research

is object detection in search and rescue (SAR) oper-

ations, surveillance, and security domains, contexts

where robust perception under uncontrollable condi-

tions is critical. The proposed solution is designed

for deployment onboard autonomous robotic plat-

forms, requiring real-time processing capabilities un-

der hardware constraints. While high-frequency in-

ference is not mandatory, achieving reliable perfor-

mance at rates around 1 Hz is sufﬁcient to support

effective operations.

Visible (RGB) images provide rich texture and

colour information, but their performance signiﬁ-

https://orcid.org/0009-0001-7717-1428

https://orcid.org/0009-0008-8213-557X

https://orcid.org/0000-0003-1117-0868

https://orcid.org/0000-0002-3045-4316

https://orcid.org/0000-0002-2245-0542

https://orcid.org/0000-0001-7811-8955

cantly degrades under poor illumination. In some

scenarios this limitation can be addressed with arti-

ﬁcial lighting; however, in the scenarios considered

here, such approach is not feasible. To ensure robust

performance under highly varying illumination con-

ditions, data fusion emerges as a promising strategy

to enhance perception. Thermal infrared images, less

sensitive to illumination changes, provide comple-

mentary information about temperature and emissiv-

ity. Speciﬁcally, the long-wave infrared spectrum of-

fers temperature measurements between 0 and 100°C,

depending on camera settings, which is particularly

suitable for pedestrian detection. However, thermal

images are sensitive to environmental changes, such

as ambient temperature and seasonal variations. As

described in (Heredia-Aguado et al., 2025), there are

corner cases—such as occlusions present in only one

modality—that further justify the integration of both

data sources.

As indicated before, visible spectrum image in-

clude rich information about textures, colour and im-

age structure, while thermal image adds light invari-

ance. Effective fusion should exploit these comple-

mentary characteristics to improve object detection in

low-light or visually degraded environments. Beyond

the domains already mentioned, multispectral fusion

has potential applications in other ﬁelds as well.

Heredia-Aguado, E., Alfaro-Pérez, M., Flores, M., Paya, L., Valiente, D. and Gil, A.

A Robust Comparative Study of Adaptative Reprojection Fusion Methods for Deep Learning Based Detection Tasks with RGB-Thermal Images.

DOI: 10.5220/0013761800003982

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

In Proceedings of the 22nd International Conference on Informatics in Control, Automation and Robotics (ICINCO 2025) - Volume 1, pages 313-320

ISBN: 978-989-758-770-2; ISSN: 2184-2809

313

Several fusion strategies exist in the literature.

This research builds upon static early fusion tech-

niques (Heredia-Aguado et al., 2025) to establish a

robust baseline for comparing the beneﬁts of various

approaches. Speciﬁcally, the fusion methods explored

in this work reduce a four-channel image (RGB+T) to

a three-channel representation, enabling the use of es-

tablished image detectors such as YOLOv8. This pa-

per extends the study of static early fusion techniques

by:

• Proposing and rigorously evaluating dynamic

fusion methods based on projections (PCA,

FA) and frequency-domain transforms (Wavelets,

Curvelets).

• Ensuring robust experimental control: using the

same dataset, network architecture, initialization,

and training parameters across all methods.

The following Section (2) details the methodology

employed, including the detection algorithm, dataset,

experimental setup, and evaluation metrics. Section 3

explains each fusion method. Section 4 presents and

discusses the detection results for each approach. Fi-

nally, Section 5 summarizes the main ﬁndings, limi-

tations, and potential future directions.

2 METHODOLOGY

The fusion algorithms that are tested in this research

combine both thermal and visible spectrum images

before they are fed to train and validate a deep learn-

ing model to perform the detection for each speciﬁc

fusion approach, as shown in Figure 1. This sec-

tion includes a detailed description of each block de-

picted in Figure 1 along with an explanation of the

relevant experimental setup and the evaluation met-

rics that will be involved in the later analysis and dis-

cussion.

For the sake of reproducibility, the source code is

openly available (https://github.com/enheragu/

yolo_test_utils/tree/v2.1.0-icinco, accessed

on 01 July 2025), allowing for complete replication of

the experiments.

2.1 Detection Algorithm

All the fusion algorithms included in this research

have been tested under the same detection network.

The idea of the network is to work as a common de-

scriptor of how each method performs. For this pur-

pose YOLOv8 (Jocher et al., 2023) have been se-

lected. YOLO is a state-of-the-art single-stage ob-

ject detection architecture that uniﬁes both localiza-

Table 1: LLVIP test and train subset characteristics.

Set Name Images Backgrounds Instances

Test 3463 0 8302

Train 12,025 2 34,135

tion and classiﬁcation within a single neural network.

Although it efﬁciently extracts multi-scale features

in three different sizes, it is a known issue that the

detection may suffer with smaller objects that dis-

appear through the convolution layers (Jiang et al.,

2022) (Diwan et al., 2023).

Although there are other detectors that could po-

tentially overcome these issues, the use case domain

in which this research is focused need a fast and low-

cost consuming solution. In terms of speed YOLOv8

has proven to be faster than the other state-of-the-art

family of solutions based on RCNN Networks such as

Faster-RCNN (Ren et al., 2017). Although there are

detection models based on transformers (Carion et al.,

2020) these models are still quite big in terms of train-

able parameters and memory consumption. YOLOv8

offers a reliable solution with a fast processing time

and low memory and trainable parameters, implying a

shorter training time and the need for smaller datasets.

2.2 The LLVIP Dataset

For this research the LLVIP dataset (Jia et al., 2021)

has been chosen. It includes up to 16 k colour-

thermal image pairs with pedestrians manually la-

belled, captured at night under low illumination con-

ditions. Images provided (1280 x 1024) have been

captured with a static pair of cameras in 26 differ-

ent locations. The dataset is already provided with

a 80-20 split for train/test subsets. As summarised

in Table 1 the images have been labelled and ﬁltered

so that all pedestrians are identiﬁed removing almost

all only-background images. This dataset seems to

be one of the most reliable datasets in terms of mul-

tiespectral images. Although bigger datasets such as

the KAIST (Hwang et al., 2015) dataset exist, they

have not the quality in terms of image pair align-

ment (Heredia-Aguado et al., 2025) and labelling as

LLVIP. Note that the fusion methods proposed are de-

pendant on the alignment of both visible and thermal

images and that problems in calibration or synchro-

nization between images can greatly impact the per-

formance of the methods.

Examples of both test and training subsets are

shown in Figure 2. As can be observed, all the images

have been taken from a high-angle perspective, which

means that most people appear relatively close to the

camera, maintaining a consistent size and avoiding

smaller instances that, as already stated, would neg-

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

314

Figure 1: Architectural design for the fusion algorithm evaluation.

atively affect the detection algorithm.

Figure 2: LLVIP examples from test and train subsets.

2.3 Experimental Setup

As already stated, the idea of this research is to estab-

lish a comparison framework between different fusion

methods. As some of them (middle fusion, late fu-

sion) imply changes in the deep-model architecture,

no transfer learning is to be used. All the models

are trained from scratch, making use of the result-

ing fused images, and based on the same initialization

weights and identical hyperparameters, including data

augmentation strategies, and initialization schemes.

This guarantees a robust and fair comparison across

the different fusion algorithms

All the training and validation tests were per-

formed using the same hardware: an NVIDIA GPU,

model GeForce RTX 4090 with 24 GB. The adapta-

tive fusion of each image is done in a 11th Gen In-

tel(R) Core(TM) i7-11700 (2.50GHz) processor.

2.4 Evaluation Metrics and

Implementation Details

To have a clearer understanding of the performance

and focusing on the use case already presented, the

performance of each method will be compared based

on the precision and recall metrics. Precision provides

a measure of the ability of the trained model to avoid

false positives while recall informs about the capacity

of the model to detect all instances without leaving

undetected ones.

The analysis also includes the standard metric

mean Average Precision (mAP) at standard IoU (In-

tersection over Union) thresholds. Fusion time is also

reported for each method, as computational efﬁciency

is critical for real-time applications.

3 IMAGE FUSION

The fusion methods that are presented cover differ-

ent alternatives of reprojecting the four channel infor-

mation into a three channel output. The ﬁrst couple

of methods are based on reprojecting the data into a

different frame based on data variation. The follow-

ing methods include fusion in the frequency domain

before data are reprojected back to the three-channel

format.

This section covers PCA and FA as methods of

dimensionality reduction applied to image fusion;

and Wavelet and Curvelet transform for image fu-

sion. Other alternatives tested such as t-SNE (Van der

Maaten and Hinton, 2008) or UMAP (McInnes et al.,

2020) may take up to 40 minutes of computation time

for each image, which is completely out of the scope

of the current research use case domain.

As already introduced, these resulting images are

then fed to the deep detection network, YOLOv8, to

A Robust Comparative Study of Adaptative Reprojection Fusion Methods for Deep Learning Based Detection Tasks with RGB-Thermal

Images

315

train and validate the resulting model.

3.1 Projection-Based Fusion

3.1.1 Principal Component Analysis (PCA)

Principal Component Analysis involves a mathemat-

ical tool that transforms a given number of corre-

lated variables into a number of uncorrelated vari-

ables. With this approach, and starting with four-

channel data, through PCA the maximum variance di-

rections are computed. Taking the most relevant com-

ponents (three components in this case), the image is

backprojected to the image space. This method has

been proposed with different variants and applications

in the ﬁeld of image processing (Kumar and Muttan,

2006) (Elmasry et al., 2020).

For this research the generic approach is followed,

for each image the data are reprojected based on the

three most relevant components and then fed to the

deep-learning algorithm.

3.1.2 Factor Analysis (FA)

Following a similar approach to the PCA tool, Factor

Analysis is another tool for dimensionality reduction

based on data variance (Joliffe and Morgan, 1992).

With this tool a set of factors are computed (similar

to components in PCA), so that the input variables

are assumed to be linear combinations of these factors

plus, for each variable, an error term. The key advan-

tage of the method is that it allows the separation of

the common variance in the data from the variance at-

tributable to error. This way, the reprojection is made

only through the factors computed based on common

variance. Although it is not a method commonly used

in image processing, we think it brings some interest-

ing approach to the problem as noise or even outliers

are not an unknown thing in image processing.

3.2 Frequency-Domain Fusion

3.2.1 Wavelet Transform Fusion

The Discrete Wavelet Transform (DWT) (Sifuzza-

man et al., 2009) is a derived technique based on the

Fourier Transform. The Fourier transform analyses a

given signal based on its frequency components, but

in doing so, it loses spatial information about the data.

Two-dimension DWT ensures maintaining spatial in-

formation (critical when analysing an image) while

focusing on the frequency analysis (Zhang, 2019).

With this approach a given image can be decomposed

into frequency components.

For each channel of the four-channel input im-

age, the DWT provides a set of frequency compo-

nents. These components are splitted into two sub-

bands: approximation coefﬁcients (cA) and detail co-

efﬁcients sub-bands. These coefﬁcients are the ones

mixed between images: RGB approximation coef-

ﬁcient is mixed with thermal approximation coefﬁ-

cient; the same applies to the detail coefﬁcients. Once

fused, an inverse transform is applied to reconstruct a

three-channel image.

The detail sub-bands, the high frequency com-

ponent of the image, capture mainly local changes,

textures and edges information, while approxima-

tion sub-bads, the low frequency components, contain

most of the general structure and spectral image infor-

mation. There are different approaches on how these

components should be combined:

• Maximum value: Between two given components

the maximum value is maintained disregarding

the other one. This approach ensures maintaining

texture and edge information, but it can include

higher noise in the resulting image.

• Average value: Although averaging both compo-

nents can diminish local changes and edges, it

maintains a smoother and cleaner image.

Both versions have been implemented into two

Wavelet fusions: averaged and max-value. For the

ﬁrst version the components of each RGB channel are

averaged with the ones from the thermal image, for

both approximation and detail sub-bands. The second

approach combines each RGB channel detail compo-

nents with the thermal components keeping the maxi-

mum value. In this case the approximation coefﬁcient

is combined following an α Blending Fusion (Oﬁr,

2023): C

Aprox

= α · C

A−RGB

+(1 − α) · C

A−T H

be-

ing alpha a relative coefﬁcient based on thermal pixel

value. The max-value method tends to preserve more

information than the average method in image fusion,

as it selects the highest intensity pixel from the input

images, ensuring that no signiﬁcant details from any

source are lost, whereas averaging can dilute or blur

important features (Patil et al., 2013) (Sahu and Sahu,

2014). Other techniques that show interesting results

are based on max-contrast (Indira, 2015) but have not

been tested as would potentially increase time con-

sumption.

3.2.2 Curvelet Transform Fusion

The problem with the Wavelet transformation is that

it focuses on point singularities, ignoring some of the

geometric properties of the structures in the image.

In addition, it does not take advantage of edge reg-

ularity (Ma and Plonka, 2010). The Curvelet trans-

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

316

form deﬁnes a different transform from the DWT, as

it performs a multi-scale and multi-directional analy-

sis that is particularly effective when representing and

compressing edge and curve structures. Not surpris-

ingly, it is widely popular for image processing solu-

tions (Starck et al., 2002).

As in the DWT fusion approach, two versions

of the algorithm have been implemented. In both,

the ﬁrst layer coefﬁcients are merged averaging both

RGB and Thermal coefﬁcients. In the last levels

(up to four levels in total) are either averaged or

maximum-ﬁltered.

In Figure 3 an example of each fusion for a given

image is presented. As it can be observed both the

FA and the Curvelet fusion provide a cleaner solution

with regards to the PCA and Wavelet fusion, better

isolating the instance and relevant information from

the image. Although visually subtle, the max-value

version of each method (DWT or Curvelet) tends to

provide a clearer image (in a similar way to the dif-

ferences between DWT and Curvelet fusion results).

It can be depicted on Figure 4d, deliberately cropped

and enlarged to pixel level, how the maximum-value

fusion tends to miss some information and produce

noisier results with regards to the average version.

(a) (b)

Figure 3: Examples in false colour representation of the re-

sult of each fusion method based on an LLVIP dataset im-

age. (a) PCA fusion. (b) FA fusion. (c) Wavelet based

fusion (averaged channels). (d) Curvelet based fusion (av-

eraged channels).

For reproducibility purposes, the source code

of both reprojection methods (https://github.

com/enheragu/yolo_test_utils/blob/v2.1.

0-icinco/src/Dataset/fusion_methods/pca_

fa_compression.py, accessed on 01 July 2025)

(a) (b)

Figure 4: Detail of the resulting image for the averaged and

max-value fusion. (a) Wavelet average fusion. (b) Wavelet

max-value fusion. (c) Curvelet average fusion. (d) Curvelet

max-value fusion.

and both frequency-domain based methods (https:

//github.com/enheragu/yolo_test_utils/

blob/v2.1.0-icinco/src/Dataset/fusion_

methods/wavelets_mdmr_compression.py,

accessed on 01 July 2025) are publicly available.

4 RESULTS

Table 2: Summary of the average computation time to per-

form the fusion for each of the methods based on the images

from the LLVIP dataset.

Fusion Method Mean (s) Std (s)

PCA Fusion 0.9630 0.2030

FA Fusion 62.9443 11.9214

Wavelet Fusion 0.8314 0.0664

Curvelet Fusion 68.8148 2.2531

As already mentioned, although mAP (mean Average

Precision) is the standard when comparing different

detection deep learning algorithms, speciﬁc use case

might require a more in depth analysis. In this re-

search precision and recall are evaluated alone, and

then both mAP50 and mAP50-95 are presented.

Figure 5 depicts a precision recall curve for each

of the methods. Note that the axis have been adjusted

to focus the plot on the area of interest.

• As it can be seen, although the FA method (the

yellow line in the graph) reaches the highest preci-

sion, the recall slightly degrades. As indicated in

the Introduction, this research is focused on SAR

A Robust Comparative Study of Adaptative Reprojection Fusion Methods for Deep Learning Based Detection Tasks with RGB-Thermal

Images

317

Table 3: Detection performance (mAP) and fusion time for each method compared to Visible, LWIR and VT method

(Heredia-Aguado et al., 2025) with thermal channel equalization. The most relevant results are highlighted in bold.

Method Precision Recall mAP50 mAP50-95 Best Epoch

Visible alone 0.871 0.799 0.870 0.487 18

LWIR alone 0.961 0.914 0.966 0.655 37

VT (No eq.) 0.946 0.900 0.955 0.640 16

VT (Thermal eq.) 0.961 0.935 0.974 0.671 60

PCA Fusion 0.9602 0.9107 0.9589 0.6416 36

FA Fusion 0.9659 0.9070 0.9582 0.6368 37

Wavelet Fusion (mean) 0.9606 0.8899 0.9543 0.6405 38

Wavelet Fusion (max) 0.9525 0.8827 0.9454 0.6289 58

Curvelet Fusion (mean) 0.9537 0.8858 0.9476 0.6066 15

Curvelet Fusion (max) 0.9612 0.8602 0.9346 0.6190 64

Figure 5: Precision recall curve for all the methods included

in this research.

operations or vigilance, areas in which missing in-

stances might be more critical than in other do-

mains. For this reason, it is always advisable to

operate at a work point with a high recall, rather

than a high precision.

• With a similar behaviour, the Curvelet with max-

value (green line in the graph) fusion also reaches

a high precision before quickly deteriorating. In

this sense, the mean-fusion version of the Curvelet

(blue line in the plot) approach although starting

with a lower precision, it is capable of maintaining

a better recall, which would make it more suitable

for SAR operations.

• On the other hand, methods like the Wavelet fu-

sion (mean) (purple line) and PCA (red line), fu-

sion demonstrate a more balanced performance,

maintaining relatively high precision even as re-

call increases. From the Table 3, PCA fusion con-

sistently achieves the best mAP50 and mAP50-95

from the methods discussed in this article.

• Finally the maximum-value version of the

Wavelet fusion has a similar behaviour than the

max-value version of the Curvelet fusion method.

Although the mean-value version of the Curvelet

fusion method has a more balanced performance

than both max-value fusion (wavelet and curvelet)

it is still quite low in performance both in preci-

sion and recall, as seen in Table 3, compared to

the other methods (mean-version of the Wavelet

fusion, PCA and FA fusions.

Although precision and recall are an important

metric, execution time should not be forgotten. In

Table 2 a summary of fusion time for each method

is presented. As it can be observed, Factor Analy-

sis and Curvelet approach are quite time consuming

methods, taking more than one minute for each im-

age, while PCA and Wavelet fusion takes less than

one second, complying with the 1 Hz requirement

mentioned in the introduction. Note that the images

in the LLVIP dataset are quite big in terms of reso-

lution, these methods might be faster under other cir-

cumstances as working with images from the KAIST

dataset, that are half the pixel size with regards to the

LLVIP dataset images.

Table 3 summarizes the detection results for all

fusion methods, as well as single-modality baselines

(RGB-only, LWIR-only), using metrics from our pre-

vious work (Heredia-Aguado et al., 2025) for refer-

ence. Based on a visual inspection of the dataset im-

ages from Figure 2 it can be already advanced that the

visible channel would have a low performance as it in-

cludes pretty poor information. On the other hand the

LWIR/Thermal images has the largest quantity of in-

formation, compared to RGT images due to the nature

of the dataset involved in the tests (daylight images

could differ). Due to the invariance in pedestrian tem-

perature and the low temperature of the background

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

318

at night, pedestrians are well deﬁned, which already

advances a good performance based on this informa-

tion. These results can be compared to the VT fusion

method (Heredia-Aguado et al., 2025), presented in

the table, and tested under the same conditions as the

rest of the methods described in this manuscript (night

condition based on LLVIP dataset). The poor colour

information that can be observed explains the per-

formance of the VT method (Heredia-Aguado et al.,

2025) as it is based on combining the intensity chan-

nel from the RGB image with the thermal channel

from the LWIR image, disregarding the colour infor-

mation. Note that this method includes an equaliza-

tion of the thermal channel, that proved to be rele-

vant in enhancing the results of the methods tested,

as the non equalized version performed poorly with

regards to the presented methods (Heredia-Aguado

et al., 2025). In terms of mAP50 and mAP50-95 both

PCA fusion and FA fusion seem to be promising, and

might beneﬁt a lot from a thermal channel equaliza-

tion.

Table 2 showed the average computation time of

each fusion method. Once the image has been fused

it is fed to the detection algorithm, YOLOv8, to be

trained on this speciﬁc image set. It is important

to review the difﬁculties the network might have to

reach the best performance, depicted in Figure 5, on

each image fusion type. In terms of training duration,

PCA, FA, and Wavelet fusion methods all achieved

similar performance within a comparable number of

epochs. The LWIR-alone approach also required a

similar training time, whereas the VT fusion method

needed almost twice as many epochs to reach its best

result. Notably, the results from both the equalized

and non-equalized versions of VT fusion indicate that

applying an equalization phase increases the training

time required for all methods, but also leads to im-

proved performance.

4.1 Discussion

Despite the sophistication of the dynamic fusion

methods, improvements with regards to previous and

simpler fusion methods are limited on LLVIP, likely

due to the dataset’s high image quality and lack of

challenging occlusions.

Notably, mean-based fusion in Wavelets outper-

formed max-value based fusion for detail coefﬁcients,

suggesting that averaging may better preserve subtle

features in detection tasks for this speciﬁc use case.

In the case of the Curvelet fusion, although the preci-

sion was improved, the recall suffered a relevant drop

in terms of recall. Contrary to some of the bibliogra-

phy already presented, mean-based fusion proved to

be more balanced for detection tasks based on RGB-

Thermal image fusion.

Although both Factor Analysis (FA) and Curvelet-

based fusion methods offer theoretically appealing

approaches for extracting and combining complemen-

tary information from RGB and thermal images, their

practical application in this study revealed signiﬁcant

practical limitations due to computational speed. In

our experiments, these methods proved to be con-

siderably slower than other fusion strategies, making

them less suitable for real-time deployment scenar-

ios, such as onboard robotic systems for search and

rescue or surveillance. The high computational cost

associated with FA and Curvelet transforms, partic-

ularly during the transformation and inverse recon-

struction stages, poses a substantial bottleneck, espe-

cially when compared to more efﬁcient methods like

PCA or deep learning-based early fusion. This trade-

off must be carefully considered when selecting fu-

sion algorithms for time-sensitive applications.

5 CONCLUSIONS

We present a robust comparative study of dynamic fu-

sion methods for RGB and thermal images, evaluated

with YOLOv8 on LLVIP dataset. While advanced

methods do not always outperform simpler baselines

in controlled settings, our work underscores the im-

portance of methodological rigour and transparent re-

porting of negative results. Such studies are essential

for advancing the state of the art and guiding future

research in multispectral fusion.

It is important to note that the presented study is

limited to well-aligned, high-quality images. Future

work should evaluate fusion methods on more chal-

lenging datasets, including occlusions, or diverse en-

vironmental conditions.

As already mentioned, it is important to evaluate

these methods with different equalization techniques,

as it proved to be relevant with early fusion methods

for the detection task (Heredia-Aguado et al., 2025).

Future work should take this into account. Although

under the conditions described, the methods were not

able to outperform LWIR alone or VT (with thermal

equalization), the results from VT without equaliza-

tion suggest that there is potential for improvement

with better fusion strategies.

ACKNOWLEDGEMENTS

This research work is part of the project funded

by ”AYUDAS A LA INVESTIGACI

ON 2025

A Robust Comparative Study of Adaptative Reprojection Fusion Methods for Deep Learning Based Detection Tasks with RGB-Thermal

Images

319

DEL VICERRECTORADO DE INVESTIGACI

Y TRANSFERENCIA” of the Miguel Hern

andez

University; the project PID2023-149575OB-I00

funded by MICIU/AEI/10.13039/501100011033 and

by FEDER, UE; and the project CIPROM/2024/8,

funded by Generalitat Valenciana, Conselleria de Ed-

ucaci

on, Cultura, Universidades y Empleo (program

PROMETEO 2025).

REFERENCES

Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov,

A., and Zagoruyko, S. (2020). End-to-end object de-

tection with transformers. In Vedaldi, A., Bischof, H.,

Brox, T., and Frahm, J.-M., editors, Computer Vision

– ECCV 2020, pages 213–229, Cham. Springer Inter-

national Publishing.

Diwan, T., Anirudh, G., and Tembhurne, J. V. (2023). Ob-

ject detection using yolo: challenges, architectural

successors, datasets and applications. Multimedia

Tools and Applications, 82(6):9243–9275.

Elmasry, S. A., Awad, W. A., and Abd El-hafeez, S. A.

(2020). Review of different image fusion techniques:

Comparative study. In Ghalwash, A. Z., El Khameesy,

N., Magdi, D. A., and Joshi, A., editors, Internet of

Things—Applications and Future, pages 41–51, Sin-

gapore. Springer Singapore.

Heredia-Aguado, E., Cabrera, J. J., Jim

enez, L. M., Va-

liente, D., and Gil, A. (2025). Static early fusion tech-

niques for visible and thermal images to enhance con-

volutional neural network detection: A performance

analysis. Remote Sensing, 17(6).

Hwang, S., Park, J., Kim, N., Choi, Y., and Kweon, I. S.

(2015). Multispectral pedestrian detection: Bench-

mark dataset and baseline. In 2015 IEEE Conference

on Computer Vision and Pattern Recognition (CVPR),

pages 1037–1045.

Indira, K. (2015). Image fusion for pet ct images using

average maximum and average contrast rules. Int J

Appl Eng Res, 10(1):673–80p.

Jia, X., Zhu, C., Li, M., Tang, W., and Zhou, W. (2021).

Llvip: A visible-infrared paired dataset for low-light

vision. In Proceedings of the IEEE/CVF International

Conf. on Computer Vision, pages 3496–3504.

Jiang, P., Ergu, D., Liu, F., Cai, Y., and Ma, B. (2022). A re-

view of yolo algorithm developments. Procedia Com-

puter Science, 199:1066–1073. The 8th International

Conference on Information Technology and Quantita-

tive Management (ITQM 2020 & 2021): Developing

Global Digital Economy after COVID-19.

Jocher, G., Chaurasia, A., and Qiu, J. (2023). YOLOv8 by

Ultralytics.

Joliffe, I. and Morgan, B. (1992). Principal component

analysis and exploratory factor analysis. Statistical

Methods in Medical Research, 1(1):69–95. PMID:

1341653.

Kumar, S. S. and Muttan, S. (2006). PCA-based image

fusion. In Shen, S. S. and Lewis, P. E., editors,

Algorithms and Technologies for Multispectral, Hy-

perspectral, and Ultraspectral Imagery XII, volume

6233, page 62331T. International Society for Optics

and Photonics, SPIE.

Ma, J. and Plonka, G. (2010). The curvelet transform.IEEE

Signal Processing Magazine, 27(2):118–133.

McInnes, L., Healy, J., and Melville, J. (2020). Umap: Uni-

form manifold approximation and projection for di-

mension reduction.

Oﬁr, N. (2023). Multispectral image fusion based on su-

per pixel segmentation. In ICASSP 2023 - 2023 IEEE

International Conf. on Acoustics, Speech and Signal

Processing (ICASSP), pages 1–5.

Patil, V., Sale, D., and Joshi, M. (2013). Image fusion meth-

ods and quality assessment parameters. Asian Journal

of Engineering and Applied Technology, 2:40–45.

Ren, S., He, K., Girshick, R., and Sun, J. (2017). Faster r-

cnn: Towards real-time object detection with region

proposal networks. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 39(6):1137–1149.

Sahu, V. and Sahu, D. (2014). Image fusion using wavelet

transform: A review. Global Journal of Computer Sci-

ence and Technology, 14(F5):21–28.

Sifuzzaman, M., Islam, R., and Ali, M. (2009). Appli-

cation of wavelet transform and its advantages com-

pared to fourier transform. Journal of Physical Sci-

ence, 13:121–134.

Starck, J.-L., Candes, E., and Donoho, D. (2002). The

curvelet transform for image denoising. IEEE Trans-

actions on Image Processing, 11(6):670–684.

Van der Maaten, L. and Hinton, G. (2008). Visualizing data

using t-SNE. Journal of machine learning research,

9(11).

Zhang, D. (2019). Wavelet Transform, pages 35–44.

Springer International Publishing, Cham.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

320