An Event Camera Simulator for Arbitrary Viewpoints Based on Neural
Radiance Fields
Diego Hern
´
andez Rodr
´
ıguez
1,2
, Motoharu Sonogashira
1,2
, Kazuya Kitano
2
, Yuki Fujimura
2
,
Takuya Funatomi
2 a
, Yasuhiro Mukaigawa
2
and Yasutomo Kawanishi
1,2 b
1
Guardian Robot Project, RIKEN, Kyoto, Japan
2
Division of Information Science, Nara Institute of Science and Technology, Nara, Japan
Keywords:
Event Camera Simulation, Neural Radiance Fields.
Abstract:
Event cameras are novel sensors that offer significant advantages over standard cameras, such as high temporal
resolution, high dynamic range, and low latency. Despite recent efforts, however, event cameras remain rela-
tively expensive and difficult to obtain. Simulators for these sensors are crucial for developing new algorithms
and mitigating accessibility issues. However, existing simulators based on a real-world video often fail to gen-
eralize to novel viewpoints or temporal resolutions, making the generation of realistic event data from a single
scene unfeasible. To address these challenges, we propose enhancing event camera simulators with neural
radiance fields (NeRFs). NeRFs can synthesize novel views of complex scenes from a low-frame-rate video
sequence, providing a powerful tool for simulating event cameras from arbitrary viewpoints. This approach
not only simplifies the simulation process but also allows for greater flexibility and realism in generating event
camera data, making the technology more accessible to researchers and developers.
1 INTRODUCTION
Event cameras represent a paradigm shift in visual
sensing technology, capturing dynamic scenes with
remarkable temporal resolution and high dynamic
range. Unlike conventional frame-based cameras,
event cameras asynchronously record changes in the
intensity of the visual field, offering a unique advan-
tage in scenarios involving fast motion or challenging
lighting conditions. Since these sensors are still rela-
tively expensive and difficult to obtain, various efforts
have been made to create simulators to facilitate their
research further.
Previous simulators aim to generate event data
from RGB video by either relying on ultra-high fram-
erates (Gehrig et al., 2020; Garc
´
ıa et al., 2016) or by
interpolation of the video sequence (Hu et al., 2021).
This comes with the drawback of not being able to
generate more data from a single video. While sim-
ulators like ESIM (Rebecq et al., 2018) attempt to
tackle this issue with the use of 3D models, generating
data that resembles a realistic scene is both time and
labor-intensive, making it unsuitable for researchers
a
https://orcid.org/0000-0001-5588-5932
b
https://orcid.org/0000-0002-3799-4550
who want to generate for their own environment.
To address this challenge of generating event data
from arbitrary viewpoints, we propose a framework
of a simulation shown in Fig. 2. The framework gen-
erates synthetic event camera data using neural radi-
ance fields (NeRFs) (Mildenhall et al., 2020), a re-
cent breakthrough in the field of computer vision that
enables the reconstruction of high-fidelity 3D scenes
from a sparse set of 2D images by leveraging neu-
ral networks to model the volumetric radiance field.
By integrating NeRF with event-based sensing prin-
ciples, we aim to create a versatile framework that
can produce realistic and diverse event camera data,
facilitating the advancement of event-based vision al-
gorithms. Notably, our method focuses on generating
event data from static scenes, allowing for the explo-
ration of how camera motion alone influences event
generation without the added complexity of dynamic
scene changes.
Our approach offers several significant advan-
tages. First, it allows for creating extensive datasets
without the need for labor-intensive data collection
processes. Second, it provides a controlled virtual
environment where various parameters can be mod-
ified to evaluate the robustness of event-based algo-
rithms. Finally, the synthetic data generated through
774
Rodríguez, D. H., Sonogashira, M., Kitano, K., Fujimura, Y., Funatomi, T., Mukaigawa, Y. and Kawanishi, Y.
An Event Camera Simulator for Arbitrary Viewpoints Based on Neural Radiance Fields.
DOI: 10.5220/0013388400003912
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
774-780
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

our method can serve as a valuable resource for train-
ing deep learning models, potentially improving their
performance in real-world applications.
The rest of this paper is organized as follows. In
section 2, we introduce some of the most important
works concerning event camera simulation and ex-
plain their working mechanism. We also quickly re-
view the formulation of neural radiance fields. In
section 3, we detail the proposed methodology for
synthesizing event camera data using NeRF and dis-
cuss the implementation and integration of these tech-
nologies. In section 4, we present experimental re-
sults demonstrating the effectiveness of our approach,
comparing them to actual event data streams and with
other video-to-event generation pipelines. Finally, in
section 5, we discuss our method’s limitations, possi-
ble extensions, and future work. By bridging the gap
between synthetic data generation and event-based
sensing, our work aims to accelerate research in event
cameras and pave the way for their broader adoption
and application.
2 RELATED WORK
2.1 Event Camera Simulation
Numerous event camera datasets and simulators have
been introduced over the years. In this section, we
review the most relevant ones and their specific ap-
plication scenarios. The number of publicly released
event camera simulators is small. While some of them
build upon previous research, they mostly tackle the
task differently.
Early simulators like (Mueggler et al., 2017) vi-
sually approximated an event stream by detecting sig-
nificant changes in luminance between two succes-
sive frames to create edge-like images that resemble
the output of an event camera. Most of these sim-
ulators did not discuss how to convert the simulated
events into realistic and accurate raw event streams.
Recent approaches like (Zhang et al., 2024) tackle
this by developing a statistics-based local-dynamics-
aware timestamp inference algorithm that enables the
smooth transition to the event stream. Other simula-
tors like (Joubert et al., 2021) attempt to physically
model the unique characteristics of the sensor and its
parameters, while methods such as (Zhu et al., 2019)
and (Hu et al., 2021) take a deep learning-based ap-
proach in order to approximate the outputs of a phys-
ical sensor. However, none of them take into account
the geometry of the scene, nor can they generate an
event stream outside the original path followed by the
camera. In order to circumvent this limitation, (Li
Figure 1: A pixel u of the intensity image I
t
in the event
generation model. A positive or negative event is generated
when the brightness change exceeds the threshold τ in a log-
arithmic scale. Represented in blue and red, respectively.
et al., 2018) and, most notably, (Rebecq et al., 2018)
leverage 3D models to render a scene in which a user-
defined camera path can be utilized to generate an
event stream. However, this approach poses the need
for detailed models in case a realistic scene is to be
simulated.
In ESIM (Rebecq et al., 2018), an output
event stream E is represented as a sequence of
e
i
= (t
i
, u
i
, p
i
), denoting brightness changes asyn-
chronously registered by an image I at time t and its
pixel location u
i
= (x
i
, y
i
) in the image, with a polar-
ity p
i
∈ {−1, 1}. The polarity of an event indicates a
positive or negative change in illumination according
to a logarithmic scale, quantized by negative and pos-
itive thresholds τ. The change in brightness between
two timestamps can be estimated by the difference of
intensity of a pixel u
i
of images at time t
i
and t
i−1
in
the logarithmic scale. This mechanism is illustrated
in Fig. 1 and formulated as follows.
p
i
=
(
−1 if τ < ∆(u
i
;t
i
)
1 if τ > ∆(u
i
;t
i
)
(1)
∆(u
i
;t
i
) = logI(u
i
;t
i
) − logI(u
i
;t
i−1
) (2)
2.2 Neural Radiance Fields
Neural radiance fields (Mildenhall et al., 2020) repre-
sent a scene utilizing a multi-layer perceptron (MLP)
F
θ
: (x, d) → (c, σ) that maps a position in 3D space
x = (x, y, z) and a 2D viewing direction d = (θ, φ) to
its corresponding directional emitted radiance, i.e., its
color c = (R, G, B) and volume density σ. From this
representation, the estimated emitted radiance
b
L at a
given pixel u can be calculated using the volume ren-
dering equation (Tagliasacchi and Mildenhall, 2022)
An Event Camera Simulator for Arbitrary Viewpoints Based on Neural Radiance Fields
775

Figure 2: Illustration of our method. We first train a neural radiance field and subsequently simulate a virtual event camera,
adding noise to the simulation.
with quadrature, as follows:
b
L(u) =
N
∑
k=1
T
k
(1 − exp(−σ
k
δ
k
))c
k
, (3)
T
k
= exp −
k−1
∑
m=1
σ
m
δ
m
!
, (4)
where σ
k
and c
k
are the volume density and the emit-
ted radiance, respectively, of a sampled position x
k
along the back-projected ray r through a pixel, which
has a direction d and an origin o at the camera cen-
ter. The sample x
k
= o + s
k
d has a distance s
k
from
the camera center and a distance of δ
k
= s
k+1
− s
k
be-
tween its adjacent sample x
k+1
.
Several advances have been made since the origi-
nal NeRF paper was first published. Neural network-
based approaches like (M
¨
uller et al., 2022) and (Chen
et al., 2022) have greatly reduced inference time and
increased 3D reconstruction quality, while methods
such as (Kerbl et al., 2023) completely forgo a neu-
ral representation and opt for a modified differentiable
point-based rendering technique. While we utilize
(M
¨
uller et al., 2022) as our rendering backbone in this
paper, it is worth noting that our method is radiance
field agnostic. Meaning that the method used to ren-
der the radiance field is interchangeable.
2.3 Event Cameras and NeRFs
Recent studies have explored the integration of event
cameras with NeRFs. Notable works such as (Klenk
et al., 2023), (Rudnev et al., 2023), and (Hwang et al.,
2023) have demonstrated promising results in con-
structing radiance field representations directly from
event camera data streams. In contrast, this paper
shifts focus from generating radiance fields to de-
riving event representations from existing radiance
fields.
This approach presents several advantages. By
leveraging the continuous and high-resolution nature
of NeRFs, it becomes possible to simulate event data
from arbitrary viewpoints and under varying condi-
tions without the need for specialized hardware. This
flexibility enables the creation of diverse datasets for
training and evaluating event-based algorithms, which
are often limited by the scarcity and cost of event
cameras.
However, simulating realistic event data from
NeRFs introduces unique challenges. Reproducing
sensor-specific noise and latency effects is essential
for generating data that closely mirrors real-world
conditions and addressing these challenges is critical
to ensuring that the simulated events are both physi-
cally plausible and useful for downstream tasks.
3 METHOD
3.1 Problem Formulation
Event cameras asynchronously detect changes in
pixel brightness, delivering high temporal resolution
and low-latency data. However, their high cost and
limited availability restrict widespread adoption. Re-
searchers rely on simulators to generate synthetic
event streams, yet existing simulators that rely on
RGB videos are limited to the viewpoints present in
the input sequence, preventing the generation of event
data from novel camera paths.
To address these challenges, we propose a NeRF-
based event camera simulator capable of generating
synthetic event data from arbitrary viewpoints.
The simulation consists of two stages: NeRF
training and event data generation. In the NeRF
training stage, the simulator accepts a set of images
I =
{
I
1
, . . . , I
A
}
. Using I, a neural radiance field F
θ
is
trained. In the event data generation stage, a camera
trajectory (a sequence of camera positions and orien-
tations) C = [c
1
, . . . , c
B
] are input to the simulator f
sim
.
Then, the simulator generates an event data stream E
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
776

along the given camera path C using the trained neural
radiance field F
θ
.
E = f
sim
(C; F
θ
) (5)
The overall process flow is illustrated in Fig. 2.
3.2 Event Data Generation by Sampling
Radiance Fields
Following the methodology behind ESIM’s event
generation from 3D models, our method approx-
imates the per-pixel value of the intensity image
logI(u
i
;t
i
) at pixel u
i
by using the trained F
θ
and a
selected camera position interpolated from the given
camera path C. For each pixel of the image at the
sampled camera position, the color of each pixel is
calculated by accumulating the contributions from all
sampled points along the ray following equation (3).
Since event cameras operate in brightness pixels, we
convert the sampled color images using the ITU-R
Recommendation BT601 for luma (Union, 2011), i.e.,
according to the formula:
Y (R, G, B) = 0.299R + 0.587G + 0.114B, (6)
with RGB channels in linear color space. This yields
the following equation:
I(u
i
;t
i
) = Y (
b
L(u
i
;t
i
)).
(7)
Generating a pair of logarithmic intensity im-
ages log(I(u
i
;t
i
)) and log(I(u
i
;t
i−1
)) based on user-
defined parameters, such as maximum number of
events per camera position, pixel refraction period,
and brightness change threshold τ.
We can then determine the number of predicted
events at a certain pixel location during that time win-
dow with the following equation:
n
i
=

|∆(u
i
;t)|
τ

. (8)
According to the number, events are generated with
polarity p
i
based on the positive or negative of
∆(u
i
;t).
3.3 Event Camera Noise
Although less studied than traditional RGB camera
noise, a data stream from an event camera normally
contains events that are not associated with changes
in intensity. These events are considered noise which
comes from two main sources: photon noise and leak-
age current (Guo and Delbruck, 2023). In low-
brightness conditions, photon noise is the most com-
mon source of noise, while leakage current domi-
nates high-brightness conditions. In some event cam-
era simulators like (Hu et al., 2021), the events that
are generated by photon noise are modeled as a Pois-
son process, in which the noise event rate linearly de-
creases with intensity. Further research on modeling
these noise sources were performed by Ruiming et.
al. (Cao et al., 2024) and we leverage their noise
model in our experiments.
4 EVALUATION
4.1 Experimental Settings
We utilize a modified version of instant NGP (M
¨
uller
et al., 2022) implemented in PyTorch as our NeRF
backbone. Each scene of the dataset was trained
for 350 epochs with an initial learning rate of 0.01
and with the Adam optimizer. We conduct our
experiments on the dataset provided by Mueggler
et al. (Mueggler et al., 2017) for our comparisons
since it contains images generated by a DAVIS sen-
sor (Brandli et al., 2014), which are used to train
the radiance field, as well as camera positions from
an external tracker, eliminating the need to use
COLMAP (Sch
¨
onberger and Frahm, 2016) for cam-
era pose estimation.
To perform our tests, we interpolate five equidis-
tant positions between each camera pose along the
initial camera path, akin to the frame interpolation
V2E does.
4.2 Evaluation Metrics
For evaluation, it is difficult to directly compare the
generated event data and the ground truth, we accu-
mulated events into an image and performed image-
level comparison.
First, an accumulation operation is performed on
both the ground truth and simulated event streams to
generate a frame representation. The accumulation
operation integrates events over time into a frame-by-
frame basis, aggregating changes captured by the sen-
sor. As shown in (Mueggler et al., 2017), a logarith-
mic intensity image log
ˆ
I(u;t) can be reconstructed
from the event stream at any point in time t by ac-
cumulating events e
i
= (t
i
, u
i
, p
i
) according to the fol-
lowing function:
log
b
I(u;t) = logI(u; 0) + γ(u;t), (9)
γ(u;t) =
∑
0<t
i
≤t
p
i
τδ(u − u
i
)δ(t −t
i
), (10)
where I(u; 0) is the rendered image at time t = 0, and
δ selects the pixel to be updated on every event (pixel
u
i
of
b
I is updated at time t
i
).
An Event Camera Simulator for Arbitrary Viewpoints Based on Neural Radiance Fields
777

(a) Ground truth
(b) Ours (No added noise)
(c) Ours (with added noise)
Figure 3: Comparison of event streams, positive and nega-
tive events are colored red and green, respectively.
We utilize a modified version of this function,
which applies a decay parameter to reduce the noise
of the generated frame. The accumulator function ap-
plies an exponential decay d(t, τ) to equation (7):
log
b
I(u;t) =log

I(u; 0)d(t, τ) + I(u;t)(1 − d(t, τ))
+
∑
0<t
i
≤t
p
i
τδ(u − u
i
)d(t −t
k
, τ)

, (11)
d(t, τ) =exp

−
t
τ

, (12)
(a) Ground truth
(b) Ours (No added noise)
(c) Ours (with added noise)
Figure 4: Zoomed-in view of a specific region from Figure
3, highlighting finer details of the event streams.
where log(I(u;0)) is the logarithm of the inten-
sity of the pixel at the previous accumulated frame,
log(I(u;t)) is a neutral potential, and the decay pa-
rameter is the time constant τ. For our experiments
we set τ = 1 × 10
−5
microseconds and log(I(u; 0)) =
0.5.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
778

(a) Ground truth (b) Ours (c) V2E (Hu et al., 2021) (d) V2CE (Zhang et al., 2024)
Figure 5: Visual comparison of accumulated frames. All frames were obtained after accumulating events according to the
process described in section 4.2.
Table 1: Comparison of PSNR (dB) values obtained in
scenes from the dataset (Mueggler et al., 2017) (higher is
better).
Scene name Ours V2E V2CE
slider 30.01 29.40 29.42
boxes 6DoF 28.32 28.06 28.15
poster 28.04 28.57 28.64
4.2.1 Qualitative Comparison
We compare frame-level results with the real event
camera stream (ground truth). We also compare the
accumulated results with V2E and V2CE to the real
event camera stream (accumulated ground truth).
4.2.2 PSNR of Accumulated Event Frames
To measure the correctness of the simulated events
quantitatively, we perform an evaluation using a Peak
Signal Noise Ratio (PSNR) basis, which is a well-
known evaluation metric of image quality (Hor
´
e and
Ziou, 2010). These PSNR comparisons are summa-
rized in Tab. 1.
4.3 Experimental Results
As demonstrated in Fig. 3, our simulator correctly ap-
proximates the positive and negative events measured
by an actual event camera. It is worth noting that due
to not including both noise and hot pixel simulation in
our experiments, some areas of the simulation appear
not to show any information registered; a zoom-in of
an extreme case is illustrated in Fig. 4.
An example of the qualitative results of the accu-
mulated images is shown in Fig. 5.
While this paper primarily focuses on the applica-
tion of radiance fields for static scene reconstruction,
it is important to note several limitations and potential
avenues for future research.
Radiance fields have the ability to reconstruct dy-
namic scenes. The NeRF backbone utilized in our
experiments did not have the capability to represent
dynamic scenes, so we left their implementation as a
task for future research.
Our simulator, by its design, does not rely on a
specific representation of radiance fields. This flexi-
bility allows for easy integration with alternative ren-
dering techniques such as Gaussian splatting (Kerbl
et al., 2023).
While our simulator demonstrates promising re-
sults in controlled environments, generalizing these
findings to real-world applications presents additional
challenges. Factors such as varying lighting condi-
tions, occlusions, and reflective surfaces can signifi-
cantly impact the performance and accuracy of radi-
ance field reconstruction.
5 CONCLUSION
In this paper, we introduced a novel method for
event camera simulation using neural radiance fields.
Our approach leverages the capabilities of NeRFs to
synthesize novel views of complex scenes, enabling
the generation of realistic and diverse event camera
data from arbitrary viewpoints. Experimental results
demonstrate that our simulator matches or outper-
forms existing methods in terms of accuracy and re-
alism, providing a valuable tool for the development
and evaluation of event-based vision algorithms. The
key contributions of this work include the integration
of NeRFs with event-based sensing principles and the
development of a versatile and efficient event camera
simulator. We believe that this method represents a
significant advancement in the field of event camera
simulation, making this technology more accessible
to researchers and developers.
REFERENCES
Brandli, C., Berner, R., Yang, M., Liu, S.-C., and Delbruck,
T. (2014). A 240 × 180 130 db 3 µs latency global
shutter spatiotemporal vision sensor. IEEE Journal of
Solid-State Circuits, 49(10):2333–2341.
An Event Camera Simulator for Arbitrary Viewpoints Based on Neural Radiance Fields
779

Cao, R., Galor, D., Kohli, A., Yates, J. L., and Waller, L.
(2024). Noise2image: Noise-enabled static scene re-
covery for event cameras.
Chen, A., Xu, Z., Geiger, A., Yu, J., and Su, H. (2022).
Tensorf: Tensorial radiance fields. In European Con-
ference on Computer Vision (ECCV).
Garc
´
ıa, G. P., Camilleri, P., Liu, Q., and Furber, S. (2016).
pydvs: An extensible, real-time dynamic vision sensor
emulator using off-the-shelf hardware. In IEEE Sym-
posium Series on Computational Intelligence, pages
1–7.
Gehrig, D., Gehrig, M., Hidalgo-Carri
´
o, J., and Scara-
muzza, D. (2020). Video to events: Recycling video
datasets for event cameras. In IEEE/CVF Conference
on Computer Vision and Pattcalern Recognition.
Guo, S. and Delbruck, T. (2023). Low cost and latency
event camera background activity denoising. IEEE
Transactions on Pattern Analysis and Machine Intel-
ligence, 45(1):785–795.
Hor
´
e, A. and Ziou, D. (2010). Image quality metrics: Psnr
vs. ssim. In In International Conference on Pattern
Recognition, pages 2366–2369.
Hu, Y., Liu, S. C., and Delbruck, T. (2021). v2e: From
video frames to realistic DVS events. In IEEE/CVF
Conference on Computer Vision and Pattern Recogni-
tion Workshops (CVPRW). IEEE.
Hwang, I., Kim, J., and Kim, Y. M. (2023). Ev-nerf:
Event based neural radiance field. In Proceedings of
the IEEE/CVF Winter Conference on Applications of
Computer Vision (WACV), pages 837–847.
Joubert, D., Marcireau, A., Ralph, N., Jolley, A., van
Schaik, A., and Cohen, G. (2021). Event camera
simulator improvements via characterized parameters.
Frontiers in Neuroscience, 15.
Kerbl, B., Kopanas, G., Leimk
¨
uhler, T., and Drettakis,
G. (2023). 3d gaussian splatting for real-time radi-
ance field rendering. ACM Transactions on Graphics,
42(4).
Klenk, S., Koestler, L., Scaramuzza, D., and Cremers, D.
(2023). E-nerf: Neural radiance fields from a moving
event camera. IEEE Robotics and Automation Letters.
Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas,
D., Ye, Q., Huang, Y., Tang, R., and Leutenegger, S.
(2018). Interiornet: Mega-scale multi-sensor photo-
realistic indoor scenes dataset. In British Machine Vi-
sion Conference.
Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T.,
Ramamoorthi, R., and Ng, R. (2020). Nerf: Repre-
senting scenes as neural radiance fields for view syn-
thesis. In ECCV.
Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., and D.
(2017). Scaramuzza: The event-camera dataset and
simulator: Event-based data for pose estimation, vi-
sual odometry, and SLAM. International Journal of
Robotics Research, 36:142–149.
M
¨
uller, T., Evans, A., Schied, C., and Keller, A. (2022).
Instant neural graphics primitives with a multiresolu-
tion hash encoding. ACM Trans. Graph., 41(4):102:1–
102:15.
Rebecq, H., Gehrig, D., and Scaramuzza, D. (2018). ESIM:
an open event camera simulator. Conference on Robot
Learning (CoRL).
Rudnev, V., Elgharib, M., Theobalt, C., and Golyanik, V.
(2023). Eventnerf: Neural radiance fields from a sin-
gle colour event camera. In Computer Vision and Pat-
tern Recognition (CVPR).
Sch
¨
onberger, J. L. and Frahm, J.-M. (2016). Structure-
from-motion revisited. In IEEE Conference on Com-
puter Vision and Pattern Recognition.
Tagliasacchi, A. and Mildenhall, B. (2022). Volume ren-
dering digest (for NeRF). arXiv:2209. 02417 [cs],
3:02417.
Union, I. T. (2011). Recommendation itu-r bt.601-7: Studio
encoding parameters of digital television for standard
4:3 and wide-screen 16:9 aspect ratios. https://www.
itu.int/rec/R-REC-BT.601.
Zhang, Z., Cui, S., Chai, K., Yu, H., Dasgupta, S., Mahbub,
U., and Rahman, T. (2024). V2ce: Video to continu-
ous events simulator.
Zhu, A. Z., Wang, Z., Khant, K., and Daniilidis, K. (2019).
Eventgan: Leveraging large scale image datasets for
event cameras. arXiv preprint arXiv:1912.01584.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
780