Enhancing Visual Odometry Estimation Performance Using Image

Enhancement Models

Hajira Saleem

1,2 a

, Reza Malekian

1,2 b

and Hussan Munir

1,2 c

Department of Computer Science and Media Technology, Malm

o University, Malm

o, 20506, Sweden

Internet of Things and People Research Centre, Malm

o University, Malm

o, 20506, Sweden

{hajira.saleem, reza.malekian, hussan.munir}@mau.se

Keywords:

Visual Odometry, Image Enhancement, Low-Light Images, Localization, Pose Estimation.

Abstract:

Visual odometry is a key component of autonomous vehicle navigation due to its cost-effectiveness and ef-

ﬁciency. However, it faces challenges in low-light conditions because it relies solely on visual features. To

mitigate this issue, various methods have been proposed, including sensor fusion with LiDAR, multi-camera

systems, and deep learning models based on optical ﬂow and geometric bundle adjustment. While these

approaches show potential, they are often computationally intensive, perform inconsistently under different

lighting conditions, and require extensive parameter tuning. This paper evaluates the impact of image en-

hancement models on visual odometry estimation in low-light scenarios. We assess odometry performance on

images processed with gamma transformation and four deep learning models: RetinexFormer, MAXIM, MIR-

Net, and KinD++. These enhanced images were tested using two odometry estimation techniques: TartanVO

and Selective VIO. Our ﬁndings highlight the importance of models that enhance odometry-speciﬁc features

rather than merely increasing image brightness. Additionally, the results suggest that improving odometry

accuracy requires image-processing models tailored to the speciﬁc needs of odometry estimation. Further-

more, since different odometry models operate on distinct principles, the same image-processing technique

may yield varying results across different models.

1 INTRODUCTION

Odometry estimation is an important process for the

navigation of autonomous robots, particularly in envi-

ronments that lack pre-existing maps. Odometry in-

volves estimating the self-motion of an autonomous

vehicle based on sensor measurements, predicting its

pose over time. Pose estimation aims to determine the

robot’s position and orientation relative to a reference

frame. Visual odometry offers several advantages, in-

cluding lower computational complexity compared to

other odometry methods. However, it tends to per-

form sub-optimally in low-light or dark conditions

(Zhao et al., 2021; Wisth et al., 2021; Lee et al.,

2024). Various enhancement techniques have been in-

vestigated to address this limitation, each presenting

unique advantages and trade-offs.

In this study, we evaluate the impact of gamma

transformation and four deep learning-based image

https://orcid.org/0000-0002-9596-2688

https://orcid.org/0000-0002-2763-8085

https://orcid.org/0000-0001-9376-9844

enhancement models—RetinexFormer (Cai et al.,

2023), MAXIM (Tu et al., 2022), MIRNet (Zamir

et al., 2022), and KinD++ (Zhang et al., 2021)—on

visual odometry estimation under low-light condi-

tions. For our experiments, we used four sequences

(01, 06, 07, and 10) from the KITTI dataset (Geiger

et al., 2012), which consists of 11 sequences of im-

ages with ground truth poses. We randomly chose

these sequences without bias towards any particular

result. Although the KITTI sequences include im-

ages with mixed lighting conditions, they do not rep-

resent extremely dark conditions, such as those in

the evening. Therefore, we artiﬁcially darkened the

KITTI images to simulate low-light conditions for our

testing.

We chose TartanVO (Wang et al., 2021) and Selec-

tive VIO (Yang et al., 2022) for the odometry estima-

tion evaluation. We selected TartanVO based on the

’its designers’ claim that it can generalize to various

environmental conditions. We chose Selective VIO

for its ability to achieve near ground truth odometry

while being resource-efﬁcient due to its lower com-

putational cost.

Saleem, H., Malekian, R. and Munir, H.

Enhancing Visual Odometry Estimation Performance Using Image Enhancement Models.

DOI: 10.5220/0012932600003822

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

In Proceedings of the 21st International Conference on Informatics in Control, Automation and Robotics (ICINCO 2024) - Volume 1, pages 293-300

ISBN: 978-989-758-717-7; ISSN: 2184-2809

293

Low-light conditions exacerbate image degrada-

tion issues such as noise and color distortion, common

in settings with limited camera quality. Simply in-

creasing brightness can worsen these issues by ampli-

fying image artifacts. Therefore, effective low-light

enhancement requires not only brightening shadows

but also reducing noise and preserving trackable fea-

tures for accurate pose estimation.

RetinexFormer (Cai et al., 2023) enhances im-

ages by decomposing them into illumination and

reﬂectance components, adjusting light and remov-

ing degradation separately. It uses an Illumination-

Guided Transformer to manage long-range dependen-

cies, outperforming 17 other methods on 13 low-light

benchmarks. Similarly, KinD++ (Zhang et al., 2021)

uses a retinex-based approach outperforming 12 other

models on seven datasets, though DUPE (Wang et al.,

2019) showed comparable results in some cases.

MAXIM (Tu et al., 2022) enhances dark regions

using a UNet-shaped framework with spatially-gated

MLPs, combining local and global visual cues. It per-

formed well on low-light enhancement tasks, though

MIRNet (Zamir et al., 2022) had a higher Peak

Signal-to-Noise Ratio but comparable Structural Sim-

ilarity Index. MIRNet, with its multi-scale informa-

tion retention and attention mechanisms, preserves

spatial details while enriching features across scales,

making it highly effective on low-light benchmarks.

Instead of aiming to perfectly restore original im-

age quality, our study focuses on improving odom-

etry and pose estimation under low-light conditions.

Therefore, we used odometry metrics like absolute

trajectory error, relative translational error, and rela-

tive rotational error for evaluating and comparing the

results in this study.

This paper makes the following contributions:

• It presents a comparative analysis of odometry

performance acheived by gamma transformation

and four state-of-the-art image enhancement mod-

els.

• It demonstrates that none of the models performed

optimally in all scenarios, highlighting the need

for a model that enhances features useful for

odometry while removing artifacts that decrease

performance.

• It identiﬁes the strengths and weaknesses of each

enhancement model in the context of odometry,

offering practical recommendations for their use

in speciﬁc scenarios.

The rest of the paper is organized as follows: Sec-

tion II presents the related works. Section III intro-

duces the methodology. Section IV presents the re-

sults and discussion. Finally, Section V concludes the

paper, summarizing the key ﬁndings and implications

of the research.

2 RELATED WORK

Visual odometry faces challenges in low-light and

blurry conditions due to haze, motion blur, and sim-

ilar factors. Various approaches, including sensor-

fusion methods with LiDAR, have been explored to

enhance performance in such conditions (Zhao et al.,

2021; Wisth et al., 2021; Lee et al., 2024). Multi-

camera systems outperform monocular setups in dark

environments Liu et al.’s (Liu et al., 2018) , but multi-

sensor odometry increases computational costs. To

mitigate these issues, recent research focuses on en-

hancing monocular odometry through position-aware

optical ﬂow and geometric bundle adjustment (Cao

et al., 2023). Despite achieving superior results in

low-light settings, these methods often struggle with

depth estimation and object proximity issues in high-

luminance and dynamic environments.

Figure 1: Methodology diagram.

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

294

Traditional algorithms like adaptive histogram

equalization have been employed to enhance low-

light images for better odometry performance (Hao

et al., 2019; Zhang et al., 2022; Gao et al., 2022).

However, these methods often fail to account for vary-

ing illumination within the same image, leading to

suboptimal results. Moreover, contrary to learning-

based approaches, their parameters must be tuned

for speciﬁc lighting conditions in most cases. Con-

versely, deep learning approaches like CycleGAN and

generative adversarial networks have been proposed

to enhance low-light images while maintaining struc-

tural consistency between frames (You et al., 2023).

Efforts have also been made to integrate low-light

capabilities directly into odometry estimation neural

networks using binary and deep descriptors (Alismail

et al., 2016), though these studies often lack general-

izability.

3 METHODOLOGY

The methodology employed in this study consists of

several key steps, organized into ﬁve functional units

(FUs), as illustrated in Figure 1.

3.1 FU1: Image Darkening

The original KITTI dataset sequences (01, 06, 07,

and 10) were predominantly recorded during daylight,

which does not represent the low-light conditions this

study aims to investigate. To address this, we applied

a gamma transformation to darken the images to sim-

ulate night-time conditions. This method effectively

adjusts the luminance through a non-linear mapping

of pixel intensities, allowing us to retain ﬁne details

while creating the desired low-light effect. The gen-

eral form of gamma transformation is expressed as:

out

= c · I

, where I

out

denotes the output pixel in-

tensity, c is a scaling constant, typically set to 1 for

simplicity, γ is the gamma correction parameter (we

used gamma value 0.3), and I

represents the input

pixel intensity, normalized to the range [0, 1].

3.2 FU2: Image Enhancement

The darkened images were then processed using

four state-of-the-art image enhancement methods:

RetinexFormer, MAXIM, KinD++, and MIRNet. The

purpose of this step was to assess the effect of these

enhancement models on the odometry estimation per-

formance. In Figure 2, one image from KITTI dataset

and its enhanced versions using image processing

methods is shown.

3.3 FU3: Odometry Estimation

We tested the enhanced images using two odometry

estimation models, TartanVO and Selective VIO, to

evaluate their performance in tracking and estimating

pose. These models were chosen for their robustness

in varying environmental conditions and their ability

to handle different image qualities.

3.4 FU4: Edge Enhancement

We observed that in RetinexFormer-enhanced images,

the color channels and image features seemed to dete-

riorate, yet they consistently demonstrated strong per-

formance in many cases. We attributed this to the en-

hancement of edges. To investigate this further, we

explored the impact of edge enhancement on odom-

etry estimation. As the next step in our research, we

applied edge enhancement techniques to the images.

Examples of images enhanced by RetinexFormer are

shown in Figure 3. To enhance edges, we ﬁrst ap-

plied a Gaussian blur with a sigma value of 2 to re-

duce noise and smooth the image. Then, we used the

Canny edge detector with threshold values of 100 and

200 to identify edges. To make the detected edges

more pronounced, we dilated them using a 3x3 ker-

nel. The edge map, initially in grayscale, was con-

verted to a three-channel image to match the original

image. Finally, we combined the original image with

the edge map by blending them with weights of 1.5

for the original image and -0.5 for the edges, result-

ing in a sharpened image with enhanced edges.

3.5 FU5: Performance Comparison

Finally, the odometry estimation performance across

different image versions was compared the using met-

rics: absolute trajectory error (ATE), relative trans-

lational error (t

rel

), and relative rotational error (r

rel

ATE assesses the global accuracy of the estimated tra-

jectory by comparing it to the ground truth, provid-

ing a single error value that summarizes the devia-

tion. t

rel

and r

rel

measure the translation and rotation

errors over speciﬁc distances or time intervals, respec-

tively, representing the local accuracy over short tra-

jectory segments. Lower values of ATE, t

rel

, and r

rel

indicate better performance. The analysis in this pa-

per involved a thorough examination of error curves,

evaluation metrics, and speed maps.

Enhancing Visual Odometry Estimation Performance Using Image Enhancement Models

295

Figure 2: Sample images enhanced from KITTI dataset using four enhancement models.

Figure 3: Images enhanced using RetinexFormer model.

4 RESULTS AND DISCUSSION

This section evaluates the impact of various image

enhancement techniques on visual odometry perfor-

mance using the TartanVO and Selective VIO tech-

niques. The discussion is organized based on the two

odometry models.

4.1 Odometry Performance with

TartanVO

Figure 4 shows trajectories produced using TartanVO

using images processed using various image process-

ing methods. Figure 5 and 7 show the results of odom-

etry estimation using TartanVO on KITTI sequences

01, 06, 07, and 10. These ﬁgures shows that MAXIM-

enhanced images generally provided the best overall

odometry performance when tested with TartanVO,

especially noticeable improvements in ATE for se-

quences 01 and 10. However, a signiﬁcant t

rel

value

for MAXIM-enhanced Seq-01 suggests it may strug-

gle with accurate translations over smaller segments.

This indicates that while MAXIM enhances global

trajectory consistency, it may not reliably estimate

smaller segment translations. Additionally, edge-

enhanced images with MAXIM show consistent per-

formance but highlight issues with Seq-01. Rotation

estimates exhibited minimal variation across images

processed with all four different methods as can be

seen in Figure 7.

The ATE results for Seq-06 and Seq-10 demon-

strate that odometry performance is negatively af-

fected in dark conditions. Conversely, improved per-

formance using dark images in Seq-01 and Seq-07,

compared to original KITTI images is due to en-

hanced image contrast. Despite the general improve-

ment with image enhancement models, maintaining

brightness consistency across the sequence is crucial

for better tracking and odometry estimation.

4.2 Odometry Performance with

Selective VIO

Figure 6 and 8 shows the results of odometry esti-

mation using Selective-VIO on KITTI sequences 01,

06, 07, and 10. MIRNet edge-enhanced images of-

fered the best performance for odometry estimation

with Selective VIO, with signiﬁcantly lower ATE val-

ues across most sequences, making it highly effective

for this model. Conversely, RetinexFormer showed

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

296

Figure 4: Odometry estimation trajectories using TartanVO.

the worst performance, especially when using edge-

enhanced images, indicating its unsuitability under

dark conditions. MAXIM-enhanced images provided

moderate improvements but lacked consistency, par-

ticularly in edge-enhanced scenarios. Overall, dark

images enhanced using MIRNet and KinD++, and

edge-enhanced versions of images enhanced using

MIRNet appeared promising for Selective VIO.

4.3 Impact of Edge Enhancement

The analysis showed that not all methods beneﬁted

from edge enhancement. While RetinexFormer dete-

riorated the visual appearance of images, it still per-

formed better in some cases than other models even

though the images processed by it appeared to keep

edges of objects and rest of the image content van-

ished. This prompted an investigation into edge en-

hancement’s effect on odometry estimation. Edge-

enhanced original images did not signiﬁcantly im-

prove accuracy with TartanVO, while they did im-

prove performance with Selective VIO, except when

the images were darkened. In sequences where the

original images were already well-lit or had a lower

contrast, edge enhancement sometimes degraded per-

formance. For example, in well-lit sequences, the ad-

ditional emphasis on edges introduced by enhance-

ment techniques like RetinexFormer occasionally led

to over-sharpening, which in turn reduced the over-

all quality of feature matching and tracking. Mor-

ever, edge enhancement method we used did not en-

sure that edge sharpening is consistently done across

the image sequence. This result highlights the impor-

tance of carefully selecting when and how to apply

edge enhancement, depending on the speciﬁc charac-

teristics of the image and the odometry model being

used. Compared to TartanVO, Selective VIO showed

more consistent beneﬁts from edge enhancement, es-

pecially when combined with MIRNet.

4.4 Speed and Lighting Variation

Rotation estimates exhibited minimal variation across

images processed with all models. However, er-

ror curves indicated a consistent trend where se-

quences with abrupt lighting changes and higher

speeds showed higher errors. This suggests that im-

age enhancement methods should focus on improv-

ing reliable feature extraction and maintaining con-

sistent lighting conditions rather than indiscriminately

brightening images.

5 CONCLUSION AND FUTURE

WORK

This study addressed the decline in odometry

performance under dark conditions by evaluat-

ing four deep learning-based image enhancement

techniques—MAXIM, MIRNet, RetinexFormer, and

KinD++—on darkened images from the KITTI

dataset. Our ﬁndings indicate that while MAXIM-

enhanced KITTI sequences generally performed well

Enhancing Visual Odometry Estimation Performance Using Image Enhancement Models

297

Figure 5: ATE comparison for TartanVO on different KITTI sequences.

Figure 6: ATE comparison for Selective-VIO on different KITTI sequences.

with TartanVO, not all sequences yielded optimal re-

sults. For Selective VIO, MIRNet, its edge-enhanced

versions, and KinD++ showed promise in improving

odometry performance. However, maintaining bright-

ness consistency across image sequences remains cru-

cial for reliable tracking and odometry estimation.

Most existing image enhancement methods are

general-purpose models that do not account for their

impact on odometry performance. As such, our fu-

ture work will focus on integrating odometry-aware

loss functions into the training of image enhancement

models. We also aim to validate these techniques in

real-time scenarios across diverse low-light datasets.

This study underscores the need for adaptive en-

hancement strategies tailored to the speciﬁc require-

ments of different odometry algorithms, particularly

in challenging lighting conditions. Given that Selec-

tive VIO and TartanVO models respond differently

to image enhancements, it is essential to test mul-

tiple methods to identify the best combination for

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

298

(a) TartanVO t

el (b) TartanVO r

Figure 7: Heatmaps of t

el and r

el of TartanVO.

(a) Selective-VIO t

el (b) Selective-VIO r

Figure 8: Heatmaps of t

el and r

el Selective VIO.

real-world applications. Future research should pri-

oritize developing robust, reliable navigation systems

for autonomous vehicles operating in low-light en-

vironments by incorporating odometry-aware train-

ing approaches. Additionally, future work should

also emphasize the importance of maintaining bright-

ness consistency across image sequences to improve

odometry estimation.

ACKNOWLEDGEMENTS

This work was supported by the Department of Com-

puter Science and Media Technology and Internet of

Things and People Research Centre at Malm

o Univer-

sity.

AUTHORSHIP CONTRIBUTION

STATEMENT

Hajira Saleem: Conceptualization, investigation, data

curation, Formal analysis, writing—original draft

preparation, Reza Malekian: Conceptualization, re-

view and editing, methodology, supervision, fund-

ing acquisition, Hussan Munir: Conceptualization, re-

view and editing, methodology, supervision.

REFERENCES

Alismail, H., Kaess, M., Browning, B., and Lucey, S.

(2016). Direct visual odometry in low light using bi-

nary descriptors. IEEE Robotics and Automation Let-

ters, 2(2):444–451.

Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., and Zhang,

Y. (2023). Retinexformer: One-stage retinex-based

transformer for low-light image enhancement. In Pro-

ceedings of the IEEE/CVF International Conference

on Computer Vision, pages 12504–12513.

Cao, Y.-J., Zhang, X.-S., Luo, F.-Y., Peng, P., Lin, C., Yang,

K.-F., and Li, Y.-J. (2023). Learning generalized vi-

sual odometry using position-aware optical ﬂow and

geometric bundle adjustment. Pattern Recognition,

136:109262.

Gao, W., Yang, G., Wang, Y., Ke, J., Zhong, X., and Chen,

L. (2022). Robust visual odometry based on image en-

hancement. In Journal of Physics: Conference Series,

volume 2402, page 012010. IOP Publishing.

Geiger, A., Lenz, P., and Urtasun, R. (2012). Are we ready

for autonomous driving? the kitti vision benchmark

suite. In 2012 IEEE conference on computer vision

and pattern recognition, pages 3354–3361. IEEE.

Enhancing Visual Odometry Estimation Performance Using Image Enhancement Models

299

Hao, L., Li, H., Zhang, Q., Hu, X., and Cheng, J.

(2019). Lmvi-slam: Robust low-light monocular

visual-inertial simultaneous localization and mapping.

In 2019 IEEE International Conference on Robotics

and Biomimetics (ROBIO), pages 272–277. IEEE.

Lee, D., Jung, M., Yang, W., and Kim, A. (2024). Lidar

odometry survey: recent advancements and remaining

challenges. Intelligent Service Robotics, pages 1–24.

Liu, P., Geppert, M., Heng, L., Sattler, T., Geiger, A., and

Pollefeys, M. (2018). Towards robust visual odometry

with a multi-camera system. In 2018 IEEE/RSJ In-

ternational Conference on Intelligent Robots and Sys-

tems (IROS), pages 1154–1161. IEEE.

Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik,

A., and Li, Y. (2022). Maxim: Multi-axis mlp for im-

age processing. In Proceedings of the IEEE/CVF con-

ference on computer vision and pattern recognition,

pages 5769–5780.

Wang, R., Zhang, Q., Fu, C.-W., Shen, X., Zheng, W.-S.,

and Jia, J. (2019). Underexposed photo enhancement

using deep illumination estimation. In Proceedings

of the IEEE/CVF conference on computer vision and

pattern recognition, pages 6849–6857.

Wang, W., Hu, Y., and Scherer, S. (2021). Tartanvo: A gen-

eralizable learning-based vo. In Conference on Robot

Learning, pages 1761–1772. PMLR.

Wisth, D., Camurri, M., Das, S., and Fallon, M. (2021).

Uniﬁed multi-modal landmark tracking for tightly

coupled lidar-visual-inertial odometry. IEEE Robotics

and Automation Letters, 6(2):1004–1011.

Yang, M., Chen, Y., and Kim, H.-S. (2022). Efﬁcient

deep visual and inertial odometry with adaptive visual

modality selection. In European Conference on Com-

puter Vision, pages 233–250. Springer.

You, D., Jung, J., and Oh, J. (2023). Enhancing low-light

images for monocular visual odometry in challenging

lighting conditions. International Journal of Control,

Automation and Systems, 21(11):3528–3539.

Zamir, S. W., Arora, A., Khan, S., Hayat, M., Khan, F. S.,

Yang, M.-H., and Shao, L. (2022). Learning enriched

features for fast image restoration and enhancement.

IEEE transactions on pattern analysis and machine

intelligence, 45(2):1934–1948.

Zhang, S., Zhi, Y., Lu, S., Lin, Z., and He, R. (2022).

Monocular vision slam research for parking environ-

ment with low light. International Journal of Automo-

tive Technology, 23(3):693–703.

Zhang, Y., Guo, X., Ma, J., Liu, W., and Zhang, J. (2021).

Beyond brightening low-light images. International

Journal of Computer Vision, 129:1013–1037.

Zhao, S., Zhang, H., Wang, P., Nogueira, L., and Scherer,

S. (2021). Super odometry: Imu-centric lidar-visual-

inertial estimator for challenging environments. In

2021 IEEE/RSJ International Conference on Intelli-

gent Robots and Systems (IROS), pages 8729–8736.

IEEE.

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

300