BevGAN: Generative Fisheye Cross-View Transformers

Rania Benaissa, Antonyo Musabini, Rachid Benmokhtar, Manikandan Bakthavatchalam

and Xavier Perrotton

Valeo, Brain Division, anSWer AI R&D Center, Cr

eteil 94000, France

Keywords:

Generative Models, Image Generation, Image-to-Image Translation, Driving Assistance Systems,

Autonomous Vehicles.

Abstract:

Current parking assistance and monitoring systems synthesize Bird Eye View (BEV) images to improve drivers

visibility. These BEV images are created using a popular perspective transform called Inverse Perspective

Mapping (IPM), which projects pixels of surround-view images captured by onboard ﬁsheye cameras onto a

ﬂat plane. However, IPM faces challenges in accurately representing objects with varying heights and seam-

lessly stitching together the projected surround-views due to its reliance on rigid geometric transformations.

To address these limitations, we present BevGAN, a novel geometry-guided Conditional Generative Adver-

sarial Networks (cGANs) model that combines multi-scale discriminators along with a transformers-based

generator that leverages ﬁsheye cameras calibration and attention-mechanisms to implicitly model geomet-

rical transformations between the views. Experimental results demonstrate that BevGAN outperforms IPM

and state-of-the-art cross-view image generation methods in terms of image ﬁdelity and quality. Compared to

IPM, we report an improvement of +6.2db on PSNR and +170% on MS-SSIM when evaluated on a synthetic

dataset depicting both parking and driving scenarios. Furthermore, the generalization ability of BevGAN on

real-world ﬁsheye images is also demonstrated through zero-shot inference.

1 INTRODUCTION

Automotive Surround-view Systems (SVS) assist ve-

hicles in navigating through unpredictable real-world

scenarios, reducing risks of accidents by making real-

time, high-conﬁdence decisions.

SVS involves four wide-angle ﬁsheye lens cam-

eras known as the surround-view cameras (SVCs),

which are already mounted in vehicles and offer a

large ﬁeld of view (up to 195°). Numerous manu-

facturers like BMW, Mercedes, Toyota and Hyundai,

integrate such systems into their around-view moni-

tors (TrueCar, nd). The market size of these systems

was estimated around USD 2734.6 million in 2022

with an expected increase to USD 43815.1 million by

2031 (Business Research Insights, nd).

Despite that, images captured by ﬁsheye cameras

exhibit strong radial distortions that are particularly

intensiﬁed with larger ﬁelds of view. It falls short in

providing an accurate and comprehensive understand-

ing of the distant environment since mapping real-

world coordinates onto perspective views alters the

objects appearance leading to hindered views and oc-

clusions between objects (see Figure 1). Moreover,

the scarcity of open datasets featuring ﬁsheye images

often deters their use in tasks beyond around view

montiors, such as perception systems.

Conversely, Bird Eye View (BEV) perception has

proven to tremendously enhance the vehicles percep-

tual capabilities by using a conﬁguration of six pin-

hole cameras mounted around the vehicle. BEV per-

ception provides rich semantic information, including

precise objects scaling and positioning, resulting in

accurate scene representations. This is particularly

pertinent in the context of parking assistance systems

where BEV images are showcased to assist the driver,

demanding high ﬁdelity and an accurate representa-

tion of the real scene.

Direct acquisition of BEV images is challeng-

ing due to the need for costly equipment (such as

drones or helicopters) to properly setup acquisition

sensors. Alternatively, current parking assistance

systems utilize an established perspective transform

called Inverse Perspective Mapping (IPM) (Mallot

et al., 1991). This transform projects pixels of

surround-view images into a ﬂat plane by computing a

homography matrix that relies on the camera calibra-

tions. IPM images are used in many perception tasks,

such as parking slots detection (Zhang et al., 2018;

Do and Choi, 2020; Wang et al., 2023) and Simulta-

neous Localization and Mapping (SLAM) tasks (Lee

et al., 2023).

One signiﬁcant drawback of IPM lies in its as-

sumption that the world is ﬂat since objects with

Benaissa, R., Musabini, A., Benmokhtar, R., Bakthavatchalam, M. and Perrotton, X.

BevGAN: Generative Fisheye Cross-View Transformers.

DOI: 10.5220/0013122200003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 2, pages 153-163

ISBN: 978-989-758-737-5; ISSN: 2184-433X

153

heights (e.g. vehicles, trees and poles) are severely

distorted and thus occlude other objects that are not

directly in the camera’s line of sight. Additionally, er-

rors in camera calibrations contribute to IPM’s failure

to seamlessly merge the projected perspective views

(see Figure 1f).

In commercial vehicles, the coverage range of the

BEV images is limited to a very short distance around

the ego-vehicle (i.e., ±5 meters) to diminish image

deformations. Consequently, IPM is primarily em-

ployed for parking applications rather than broader

driving scenarios which limits the drivers visibility of

their surroundings and prevents them from having a

satisfactory user experience (Musabini et al., 2021).

This work aims to overcome these challenges and

improve the driver experience across both parking and

driving scenarios. To the best of our knowledge, this

is the ﬁrst attempt to generate BEV images using Gen-

erative Adversarial Networks (GAN) by exploiting

BEV features that are directly obtained from distorted

surround-view ﬁsheye images. Our main contribu-

tions are :

• A novel geometry-guided Conditional Generative

Adversarial Networks (cGANs) model referred to

as BevGAN. The proposed architecture leverages

robust mechanisms for features extraction and in-

herently acquires mappings from individual cam-

era views to infer BEV features representation.

The latter are then transformed into BEV images

all in a GAN-like framework.

• BevGAN generates BEV images of signiﬁcantly

higher quality compared to IPM and generative

state-of-the-art methods. Furthermore, BevGAN

generation capabilities span across diverse driv-

ing scenarios, while the BEV coverage range is

expanded to a distance of ±12m around the ego-

vehicle.

2 RELATED WORK

2.1 Vision-Centric BEV Perception

Vision-centric BEV perception focuses on transform-

ing one or multiple view inputs into BEV represen-

tations that are subsequently used to perform crucial

downstream perception tasks (i.e. objects detection,

lanes and maps segmentation, etc...). Current liter-

ature is divided into two main streams: geometry-

based and network-based transformations (May et al.,

2022).

Geometry-based transformations explicitly lever-

ages the geometric properties of physical cameras to

(a) Front. (b) Rear. (c) Left. (d) Right.

(e) Ground truth. (f) IPM. (g) BevGAN.

Figure 1: We propose a geometry-guided generative ad-

versarial framework (BevGAN) that generates a BEV im-

age from four surround-view ﬁsheye images. Compared to

IPM(1f), BevGAN fuses surround-view images into a uni-

ﬁed BEV grid where the positioning of each object in the

scene is known through a learnt transform. In (1g), no dis-

tortions are observed and the objects appearances are pre-

served thanks to GANs capabilities in generating unseen ar-

eas, resulting on a scene that is more faithful to the ground

truth (1e).

establish a natural projection relationship between the

views. A pioneering work is IPM, as introduced in

Section 1. It efﬁciently transforms the views through

a homography matrix that is derived from intrinsic

and extrinsic parameters of the camera. However,

it falls short in meeting the intricacies of real-world

driving as it relies on a rigid ﬂat plane assumption that

causes noticeable distortions on objects lying above

the ﬂat plane (for example vehicles, buildings, pedes-

trians). Despite other attempts to reduce the distor-

tions by performing IPM on objects footprint segmen-

tation maps (Loukkal et al., 2021; Can et al., 2020)

and feature maps (Reiher et al., 2020) to align with

the ﬂat ground assumption, it is still present.

Network-based methods implicitly incorporate

camera geometry using neural networks to learn a

cross-view mapping (Yang et al., 2021; Zou et al.,

2023). In particular, transformers-based networks are

widely used in dense and sparse perception tasks for

their expressiveness (May et al., 2022). These meth-

ods utilize cross-attention between the BEV queries

and the input image features to transform the views

(May et al., 2022; Zhou and Krahenbuhl, 2022; Ji-

achen et al., 2022; Bartoccioni et al., 2022; Yang

et al., 2023).

However, capitalizing on the rich BEV representa-

tions to generate BEV images remains a nascent ﬁeld

of research as we report only one work (Gieruc et al.,

2024) that exploits TriPlane representations to gener-

ate BEV images.

Moreover, only few prior works consider ﬁsheye

camera geometry (Samani et al., 2023) (Musabini

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

154

et al., 2024) or use a combination of both ﬁsheye and

pinhole cameras (Pham et al., 2024).

2.2 Generative Adversarial Networks

GANs were introduced to synthesize new images

from a complex, high-dimensional training distribu-

tion (Goodfellow et al., 2020). It consists of two net-

works trained adversarially : a generator G that learns

to generate photo-realistic images from a noise vector

and a discriminator D that learns to distinguish be-

tween the generated and ground truth images.

Many variants of GANs use novel networks ar-

chitectures and loss functions to leverage speciﬁc

tasks like super-resolution (Ledig et al., 2017), text-

to-image translation (Zhang et al., 2017) or images

editing (Pan et al., 2023).

Among Conditional Generative Adversarial Net-

works (cGANs) applications, image-to-image trans-

lation (also called cross-domain image translation)

achieved prominent success in mapping images of a

source domain X to a target domain Y . Early endeav-

ors (Zhu et al., 2017; Isola et al., 2017; Wang et al.,

2018) concentrated on transforming aligned source

and target domains, where geometrical transforma-

tions between the domains views are insigniﬁcant.

Recent approaches (Jain et al., 2021; Zhu et al.,

2018) exploit semantic segmentation or geometry in-

formation (such as depth or homography estimation)

to handle unaligned domains. BridgeGAN (Zhu et al.,

2018) generates a BEV image of a driving scene from

a single frontal view image by incorporating the ho-

mography image as an intermediate view to a multi-

GAN framework designed such that frontal, homog-

raphy and bird eye views share the same feature rep-

resentation. However, this method highly depends on

homography view and distortions are still present. A

video-to-video translation model (Jain et al., 2021)

was also introduced to generate BEV driving se-

quences from egocentric RGB videos using an addi-

tional branch that estimates the optical ﬂow map to

ensure temporal consistency between the generated

frames. However, this method produces blurrier im-

ages over time due to accumulated generations errors

and some inconsistencies related to global structure

of the scene are also reported.

Shifting to semantic-guided approaches, the gen-

erator in (Regmi and Borji, 2019; Wu et al., 2023)

is forked to produce both target-view images and

segmentation maps to learn rich semantic features

through the optimization of alignment losses. In (Ren

et al., 2021; Tang et al., 2020a; Tang et al., 2019;

Ren et al., 2023), the generator is conditioned by

real target-view segmentation maps to learn shared

features that progressively align domains. Over-

all, semantic-guided frameworks mostly depend on

target-view segmentation maps at inference time or

have heavy, complex architectures that hardly learn

efﬁcient mappings between the views.

To date, none of the existing methods handle ﬁsh-

eye images and multiple surround-view images fu-

sion. In this work, we formulate our task as a cross-

view image-to-image translation task and propose a

geometry-based model that involves four surround-

view ﬁsheye images (i.e. front, rear, left and right

ﬁsheye images). Instead of relying on hard geometry

transforms, it learns an estimate of the 3D objects lo-

cations in the scene to construct pertinent BEV map-

pings. The generative capabilities of our model al-

low to translate these mappings into a BEV image that

faithfully represents the scene.

3 METHODOLOGY

3.1 The Proposed Cross-View

Transformers-Based GAN

Our goal is to generate a high-ﬁdelity BEV image

considering a set of four surround-view ﬁsheye im-

ages (I

∈ R

W ×H×3

)

k=1

each having a corresponding

camera intrinsics matrix K

∈ R

3×3

, extrinsic rotation

matrix R

∈ R

3×3

and translation vector t

∈ R

rel-

ative to the center of the ego-vehicle.

To accomplish that, we introduce a novel

geometry-guided cross-view image-to-image transla-

tion GAN architecture, referred to as BevGAN. Using

cameras calibration and attention mechanisms, it im-

plicitly models geometrical transformations between

the views to produce rich BEV features representa-

tion.

BevGAN overall architecture is presented in Fig-

ure 2. It comprises multi-scale discriminators and

a transformers-based generator designed with an en-

coder and decoder built upon cross-view transformer

(Zhou and Krahenbuhl, 2022) and Pix2PixHD (Wang

et al., 2018) frameworks respectively.

The key elements of BevGAN are outlined in the

following sections.

3.1.1 BEV Grid Construction

BEV grid representations are constructed by adapting

the dense query-based encoder of Cross-view Trans-

formers (Zhou and Krahenbuhl, 2022) to take input

images acquired from four ﬁsheye cameras, instead of

the original six pinhole cameras.

BevGAN: Generative Fisheye Cross-View Transformers

155

Fisheye-camera-aware

Positional Encodings

Generated image

Ground truth image

Image

Generation Head

Multi-scale features

B.E.V grid construction

B.E.V queries

Cross-attention layers

Fisheye images

Features

Extractor

Multi-scale discriminators

Flatten

Fake

Real

Fc + Softmax

Flatten

Fc + Softmax

Fake

Real

Figure 2: BevGAN architecture overview. Image features across multiple resolutions are extracted from surround-view

ﬁsheye images. Combined with Positional encodings (see Figure 3) obtained from cameras position and intrinsics, the BEV

grid is constructed through a serie of cross-attention layers and passed to the decoder to generate the ﬁnal BEV image. The

BEV image along with ground truth image are given to multi-scale discriminators for the discrimination process.

First, a features extractor builds up multi-scale

features φ

for each input image (see Figure 2) (Zhou

and Krahenbuhl, 2022). Two different resolutions

are considered, each processed independently to be

passed into a cross-view attention mechanism. The

latter matchs up a BEV representation with input im-

ages features by implicitly modeling the geometry

transformation between the views.

Starting from the lowest resolution, learnt BEV

embeddings are reﬁned through the projection of

higher resolutions to better estimate the 3D location

of each object in the scene.

3.1.2 Fisheye-Camera-Aware Positional

Encoding

For each image coordinate x

(I)

, the angle of incidence

k,i

emanating from each ﬁsheye camera lens is com-

puted based on its distance r

k,i

from the principal

point (center of projection) and the radial distortions

coefﬁcients of each camera (c

k,l

)

l=1

k,i

= c

k,1

k,i

+ c

k,2

k,i

+ c

k,3

k,i

+ c

k,4

k,i

(1)

Equation 1 depicts the relationship between the

angle of incidence α

k,i

and the distance r

k,i

follow-

ing the Polynomial FishEye Transform (PEFT). The

computed polynomial root, α

k,i

, encodes the depth in-

formation of a new world coordinate. The latter is

multiplied by the inverse of the camera extrinsic rota-

tion matrix R

and translation vector t

to obtain the

direction vector d

k,i

which represents a ﬁsheye cam-

era positional encoding of the image coordinate x

(I)

Figure 3 illustrates a multi-view 3D visualization of

the ﬁsheye camera positional encodings derived from

a single-frame scene.

(a) Side view. (b) Front view. (c) Top view.

Figure 3: A multi-view 3D visualization of ﬁsheye-camera

positional encodings derived from a single-frame scene.

The vehicle’s center is positioned at pixel coordinates (0, 0)

and the z-axis depicts height above ground level. The po-

sitional encoding of each camera is represented by a color

: red for front camera, yellow for rear camera, blue for left

camera and green for right camera.

The direction vector d

k,i

is encoded, using an MLP

shared across all k views, into a D-dimensional posi-

tional embedding δ

k,i

∈ R

128

3.1.3 Image Generation Head

The proposed decoder is constructed upon the global

generator of Pix2PixHD (Wang et al., 2018) which

has proven to be effective for high-resolution image-

to-image translation and style transfer tasks (Johnson

et al., 2016). It consists of a sequence of 9 residual

blocks followed by a transposed convolutional back-

end (which is composed of 3 blocks of 3 × 3 convo-

lution of stride

, instance normalization and ReLU

layers).

3.1.4 Multi-Scale Discriminators

In order to generate coherent scene representation

with ﬁner details, it is necessary to design a discrimi-

nator that has a large receptive ﬁeld. Inspired by pre-

vious works (Wang et al., 2018; Tang et al., 2020b),

we adopt multi-scale discriminators which comprise

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

156

three identical convolutional PatchGAN classiﬁers

(Isola et al., 2017) that operate at different scales. For

each discriminator D

, k = {1, 2, 3}, we downsample

input images by a factor of 2

k−1

3.2 Optimization Scheme

Three optimization loss functions are considered in

BevGAN : i) a least-square adversarial loss L

lsgan

(Mao et al., 2017) , ii) a feature matching loss L

(Isola et al., 2017) and iii) a VGG perceptual loss

V GG

(Isola et al., 2017). The generator G and dis-

criminators D

, k = {1, 2, 3} are alternately optimized

according to the following optimization problem :

BevGAN

= arg min

G,D

∑

k=1

lsgan

(G, D

λ(

∑

k=1

(G, D

) + L

V GG

(G))

(2)

where λ controls the importance given to the visual

losses L

and L

V GG

4 EXPERIMENTS

4.1 Parallel Domain Dataset

The dataset used in this work was procedurally gen-

erated using the synthetic data platform developed at

Parallel domain (Parallel Domain Plateform, nd). To

ensure visual diversity in the images, three different

scenario types were generated:

1. Highway scenario

2. Urban scenario with high pedestrian density

3. Parking scenarios with the following parking slot

variations :

• Angled parking slots

• Parallel parking slots

• Perpendicular slots

Each scenario contains high ﬁdelity renderings

(and all the associated annotations) for four surround-

view cameras and a BEV camera viewing the scene

orthogonally as illustrated in Figure 4.

4.2 Evaluation Metrics

Drawing from previous works (Wu et al., 2023; Ren

et al., 2023; Tang et al., 2020a; Regmi and Borji,

2019; Zhu et al., 2018; Ren et al., 2021), the follow-

ing quantitative metrics have been selected to assess

the quality of the generated images.

(a) Highway. (b) Urban. (c) Angled.

(d) Parallel. (e) Perpendicular.

Figure 4: Illustration of BEV samples from each scene type

of Parallel Domain dataset.

• Peak Signal-to-Noise Ratio (PSNR) operates at

the pixel-level. It is deﬁned as the ratio between

the maximum of a signal f and the power of cor-

rupting noise that affects the ﬁdelity of its repre-

sentation.

• Multi-Scale Structural SIMilarity (MS-

SSIM)(Wang et al., 2003) value ranges from 0

to 1. MS-SSIM compares two images patches at

multiple scales based on their luminance, contrast

and structure.

• Sharpness Difference (SD) measures the similar-

ity between real and generated images in terms

of sharpness by computing the difference between

their gradients.

• Learned Perceptual Image Patch Similarity

(LPIPS) computes the difference between the

feature maps patches of the real and gener-

ated images extracted at different layers of a

pre-trained model (a VGG-network, AlexNet or

SqueezeNet...) using L1-distance.

• Frechet Inception Distance (FID)(Heusel et al.,

2017) uses frechet distance to measure differences

in density between two distributions at the deepest

layers of an Inception-v3 classiﬁer.

4.3 Implementation Details

In all experiments, we use an EfﬁcientNet-V2 Medium

(Tan and Le, 2021) as the features extractor and set

the weight λ = 10 (see Equation 2). Other networks

parameters remain unchanged. BevGAN is trained

on a 40G NVIDIA A100 GPU for 80 epochs follow-

ing the splitting scheme established in Table 1. We

use AdamW optimizer and the one-cycle learning rate

scheduler (Smith and Topin, 2018) with a learning

rate of 10

−3

for both generator and discriminators.

BevGAN: Generative Fisheye Cross-View Transformers

157

Table 1: Training and test sets splitting scheme for Parallel Domain Dataset.

Scenes Highway Urban

Parking

Angled Parallel Perpendicular

Training set images 5350 6850 6720 8625 5350

Test set images 650 650 745 975 650

Total 6000 7500 7465 9600 6000

The batch size is set to 9 which is the largest size that

the model can accommodate within the GPU capacity.

Surround-view ﬁsheye images are resized to

640 × 528 and BEV images are cropped and re-

sized such that it covers an area of ±12m around the

ego-vehicle with a corresponding pixels resolution of

200 × 200 for a BEV grid size of 25 × 25.

4.4 Results

4.4.1 State-of-the-Art Comparison

The proposed BevGAN is compared to three meth-

ods : IPM, Pix2PixHD and PanoGAN. IPM is the

cutting-edge method deployed in parking assistance

and monitoring systems to transform multiple inputs

into a single BEV image.

Pix2PixHD originally generates high-resolution

images from a single input of a different domain (for

example, edges-to-image or labels-to-image transla-

tion (Wang et al., 2018)). Because BevGAN employs

similar decoder and discriminators, contrasting the

two models gives an assessment of the transformer-

based encoder ability to construct a detailed BEV

grid.

On the other hand, PanoGAN is the state-of-the-

art method for cross-view image-to-image translation.

It is designed to handle more signiﬁcant geometrical

transformations since it synthesizes a panorama im-

age from a single BEV image.

Pix2PixHD and Panogan are adapted and trained

on Parallel Domain dataset following the training

scheme in 1, using the same hyper-parameters as set

in their respective papers. Both models initially take

one single input image. To accommodate to our task,

the four ﬁsheye images were arranged in a 2 by 2 grid

then fed to the models as a single input.

Quantitative Evaluation. The quantitative evalua-

tion results obtained on Parallel Domain dataset are

presented in Table 2. It is noticeable that BevGAN

achieves the best scores on all presented metrics.

In comparison to IPM, an improvement of +6.2db

on PSNR, +170% on MS-SSIM and +22% on SD

is observed. Moreover, BevGAN outperforms both

Multi-input PanoGAN and Multi-input Pix2PixHD

by a large margin as we report an improvement of

+2.66db and +1.87db on PSNR and +35% and

+20% on MS-SSIM respectively. These results in-

dicate that BevGAN is able to produce BEV im-

ages with higher quality than other leading meth-

ods. Scores achieved on high-level evaluation metrics

(LPIPS and FID) demonstrate the expressiveness of

the BEV representations obtained with BevGAN as it

allows the generation of diverse, high-quality images.

Table 3 presents quantitative evaluation of Bev-

GAN on Parallel Domain dataset for each scene type.

It is worth mentioning that BevGAN performs better

on highway scenes in contrast to urban and parking

scenes. In fact, parking and urban scenes simulate

more complex and detailed environments, featuring

high pedestrians density in urban areas and a variety

of parking slot types and markings in parking scenar-

ios (see Figure 4). Moreover, only a small number of

scenes include challenging conditions (like night and

snow scenes). Plus, these scenes are predominant in

parking scenarios making it difﬁcult for the model to

effectively learn their characteristics.

Qualitative Evaluation. Figure 5 illustrates quali-

tative results obtained with leading methods on vari-

ous scenes of Parallel Domain dataset. It is observ-

able that BevGAN produces more realistic BEV im-

ages across all scenarios closely matching the corre-

sponding ground truth images. Compared to IPM, all

generative methods provide distortion-free represen-

tations of the generated objects. Although, BevGAN

is able to identify all objects in the scene and repre-

sents them more accurately in terms of shape, color,

and sharpness.This holds true even in complex sce-

narios (see row 2 and 5) where textured objects with

varied sizes (vehicles, pedestrians, etc...) and various

road and parking slots markings, are present. Further-

more, BevGAN synthesizes a scene where invisible

regions (ie. regions that are originally invisible in

surround-view images) are consistent with respect to

the visible parts of the scene (see the vehicles roofs,

vehicles occluded by the blue vehicle in row 5).

4.4.2 Ablation Study

An ablation study was conducted on Parallel Domain

dataset to assess the effectiveness of the proposed

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

158

Table 2: Quantitative evaluation of BevGAN against state-of-the-art methods on Parallel Domain Dataset.

Method PSNR (↑) MS-SSIM (↑) LPIPS (↓) SD (↓) FID (↓)

IPM 17.14 0.30 0.47 0.22 249.46

Multi-input PanoGAN 20.68 0.60 0.40 0.23 176.80

Multi-input Pix2PixHD 21.47 0.67 0.29 0.20 114.33

BevGAN (ours) 23.34 0.81 0.17 0.18 52.7

Table 3: Quantitative evaluation of BevGAN on Parallel Domain Dataset based on scenes type.

Scenes PSNR (↑) MS-SSIM (↑) LPIPS (↓) SD (↓) FID (↓)

Urban 23.11 0.8 0.18 0.23 89.29

Highway 25.54 0.85 0.13 0.13 42.00

Parking

Angled 23.5 0.81 0.19 0.20 89.75

Parallel 23.26 0.83 0.19 0.14 113.87

Perpendicular 21.75 0.82 0.17 0.13 116.30

Table 4: Ablation study of the proposed BevGAN on Parallel Domain Dataset.

Variant Description PSNR (↑) MS-SSIM (↑) LPIPS (↓) SD (↓) FID (↓)

A generator + VGG 18.76 0.67 0.32 0.18 108.82

B generator + discriminator + VGG 21.5 0.72 0.23 0.20 67.20

C generator + discriminator + FM 22.65 0.77 0.22 0.19 83.12

BevGAN generator + discriminator + all losses 23.34 0.81 0.17 0.18 52.7

BevGAN model. Table 4 provides a quantitative com-

parison of various BevGAN variants. Variant A in-

cludes the generator only, which is trained in a super-

vised manner using input ﬁsheye images and corre-

sponding BEV images. In this conﬁguration, the op-

timization focuses solely on minimizing the VGG loss

since that both adversarial and FM losses depend on

the discriminator (see Equation 2). Compared to this

variant, BevGAN demonstrates a signiﬁcant increase

in performances on all evaluation metrics with a dif-

ference of 4.58db in PSNR and a increase of +21%

in MS-SSIM. These results conﬁrm that the use of a

generative model is more adequate for the task.

The two remaining variants (variants B and C) in-

clude both the generator and discriminator. They eval-

uate the impact of each visual loss on the quality of

the generated images by employing only one visual

loss at a time. Results from Variant B indicate that

adding the FM loss substantially improves the perfor-

mances. Although adding VGG loss also improves

the results, its effect is less pronounced compared to

the feature matching loss, as observed from evalua-

tion results of Variant C.

4.4.3 Generalization Ability

Given the non-existence of a publicly available

dataset containing paired real-world surround-view

ﬁsheye and corresponding BEV images, all prior ex-

periments were conducted on Parallel Domain’s syn-

thetic dataset. When transitioning to real-world im-

ages, the current implementation of BevGAN shows

poor generalization capabilities (see Figure 6). Since

that direct training or ﬁne-tuning on a real dataset is

not feasible, we chose to retrain our model using the

following data augmentations to improve its general-

ization capabilities.

• Image-level data augmentations : image dropout

with a probability p ∈ [0.0, 0.01], sharpening,

change in brightness and hue channels.

• Roll rotations of one degree per camera.

• BEV Flip Left/Right and rotations of ﬁxed de-

grees (90

◦

, 180

◦

and 270

◦

) on yaw axis with a

probability of 0.8.

Table 5 provides a comparative evaluation be-

tween BevGAN and BevGAN

†

, the variant trained us-

ing previously cited augmentations. Despite that Bev-

GAN outperforms BevGAN

†

showing a difference of

0.64db in PSNR and 20% in MS-SSIM, BevGAN

†

demonstrates superior generalization capabilities, as

highlighted in Figure 6. Additional zero-shot in-

ferences of IPM and BevGAN

†

on our internal real

dataset are illustrated in Figure 7. The generated im-

ages demonstrate that BevGAN

†

accurately locates

vehicles and parking slots in the scene. Neverthe-

less, we report some distortions and a drop in tex-

tural quality as the captured distributions depend on

the synthetic dataset on which the model was origi-

BevGAN: Generative Fisheye Cross-View Transformers

159

(a) GT. (b) IPM. (c) PanoGAN. (d) Pix2PixHD. (e) BevGAN.

Figure 5: Qualitative comparison between different methods on Parallel Domain dataset. From left to right : Ground Truth,

IPM, Multi-input PanoGAN, Multi-input Pix2PixHD and BevGAN (ours). The proposed BevGAN generates more realistic

results with ﬁner details on all scenes settings compared to other state-of-the-art methods.

Table 5: Quantitative evaluation of BevGAN against

BevGAN

†

on Parallel Domain Dataset.

Method PSNR (↑)MS-SSIM (↑)LPIPS (↓)SD (↓)FID (↓)

BevGAN 23.34 0.81 0.17 0.18 52.7

BevGAN

†

22.7 0.76 0.23 0.2 70

nally trained. That being said, BevGAN

†

struggles to

reproduce objects not encountered during the training

phase, such as buildings and road markings, which are

absent from Parallel Domain Dataset.

5 CONCLUSION

In this work, we introduce BevGAN : a novel

geometry-guided cGANs model designed to generate

a BEV image from a set of four surround-view ﬁsh-

eye images. BevGAN integrates two key elements :

a cross-view transformer-based generator and multi-

scale discriminators. The generator operates across

multiple scales and leverages ﬁsheye-camera-aware

positional embeddings to generate high-quality BEV

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

160

(a) Fisheye. (b) IPM.

†

Figure 6: Example of zero-shot inference of BevGAN on

real-world ﬁsheye images.

images that faithfully represent the scene.

Experiments conducted on a synthetic dataset

demonstrate that the proposed BevGAN outperforms

leading methods used in parking assistance and vision

monitoring systems, as well as state-of-the-art GAN

approaches for cross-view generation. Moreover, ex-

periments show that with just a few added augmenta-

tion strategies, BevGAN can effectively generalize to

images acquired from real cameras.

BevGAN introduces promising performances for

practical integration into advanced around-view sys-

tems for real-world vehicles. Notably, our method

expands the covered area around the ego-vehicle to

a range of ±12m, which is a signiﬁcant improvement

compared to the ±5m coverage range offered by cur-

rent systems.

Future works will focus on exploring novel view

representations like Tri-Perspective View (TPV) rep-

resentation (Huang et al., 2023) (Gieruc et al., 2024)

for a robust description of the 3D scene. Additionally,

steps will be taken to collect a more diverse dataset

encompassing a larger range of driving and parking

scenarios.

REFERENCES

Bartoccioni, F., Zablocki, E., Bursuc, A., Perez, P., Cord,

M., and Alahari, K. (2022). Lara: Latents and rays for

multi-camera bird’s-eye-view semantic segmentation.

In 6th Annual Conference on Robot Learning.

Business Research Insights (n.d.). Business research in-

sights. Accessed: Febuary 2024.

(a) Fisheye images. (b) IPM. (c) BevGAN

†

Figure 7: Qualitative comparison between IPM and aug-

mented BevGAN on real-world ﬁsheye images.

Can, Y. B., Liniger, A., Unal, O., Paudel, D. P., and Gool,

L. V. (2020). Understanding bird’s-eye view of road

semantics using an onboard camera. IEEE Robotics

and Automation Letters, 7:3302–3309.

Do, H. and Choi, J. Y. (2020). Context-based parking

BevGAN: Generative Fisheye Cross-View Transformers

161

slot detection with a realistic dataset. IEEE Access,

8:171551–171559.

Gieruc, T., K

astingsch

afer, M., Bernhard, S., and Salzmann,

M. (2024). 6img-to-3d: Few-image large-scale out-

door driving scene reconstruction.

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., Courville, A., and Ben-

gio, Y. (2020). Generative adversarial networks.

Commun. ACM, 63(11):139–144.

Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and

Hochreiter, S. (2017). Gans trained by a two time-

scale update rule converge to a local nash equilibrium.

In Proceedings of the 31st International Conference

on Neural Information Processing Systems, NIPS’17,

page 6629–6640, Red Hook, NY, USA. Curran Asso-

ciates Inc.

Huang, Y., Zheng, W., Zhang, Y., Zhou, J., and Lu, J.

(2023). Tri-perspective view for vision-based 3d se-

mantic occupancy prediction.

Isola, P., Zhu, J.-Y., Zhou, T., and Efros, A. A. (2017).

Image-to-image translation with conditional adversar-

ial networks. CVPR.

Jain, V., Wu, Q., Grover, S., Sidana, K., Chaudhary, D.-G.,

Myint, S., and Hua, Q. (2021). Generating bird’s eye

view from egocentric rgb videos. Wireless Communi-

cations and Mobile Computing, 2021:1–11.

Jiachen, L., Zheyuan, Z., Xiatian, Z., Hang, X., and Li, Z.

(2022). Learning ego 3d representation as ray tracing.

Johnson, J., Alahi, A., and Fei-Fei, L. (2016). Perceptual

losses for real-time style transfer and super-resolution.

In Leibe, B., Matas, J., Sebe, N., and Welling, M.,

editors, Computer Vision – ECCV 2016, pages 694–

711, Cham. Springer International Publishing.

Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunning-

ham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J.,

Wang, Z., and Shi, W. (2017). Photo-realistic single

image super-resolution using a generative adversarial

network. In 2017 IEEE Conference on Computer

Vision and Pattern Recognition (CVPR), pages 105–

114, Los Alamitos, CA, USA. IEEE Computer Soci-

ety.

Lee, Y., Kim, M., Ahn, J., and Park, J. (2023). Accurate

visual simultaneous localization and mapping (slam)

against around view monitor (avm) distortion error us-

ing weighted generalized iterative closest point (gicp).

Sensors, 23(18).

Loukkal, A., Grandvalet, Y., Drummond, T., and Li, Y.

(2021). Driving among ﬂatmobiles: Bird-eye-view

occupancy grids from a monocular camera for holistic

trajectory planning. In 2021 IEEE Winter Conference

on Applications of Computer Vision (WACV), pages

51–60.

Mallot, H., B

ulthoff, H., J.J., L., and S, B. (1991). Inverse

perspective mapping simpliﬁes optical ﬂow computa-

tion and obstacle detection. Biological cybernetics,

64:177–85.

Mao, X., Li, Q., Xie, H., Lau, R. Y. K., Wang, Z., and Smol-

ley, S. P. (2017). Least squares generative adversarial

networks. In 2017 IEEE International Conference

on Computer Vision (ICCV), pages 2813–2821, Los

Alamitos, CA, USA. IEEE Computer Society.

May, Y., Wangy, T., Baiy, X., Yang, H., Hou, Y., Wang, Y.,

Qiao, Y., Yang, R., Manocha, D., and Zhu, X. (2022).

Vision-centric bev perception: A survey.

Musabini, A., Bozbayir, E., Marcasuzaa, H., and Ram

ırez,

O. A. I. (2021). Park4u mate: Context-aware digi-

tal assistant for personalized autonomous parking. In

2021 IEEE Intelligent Vehicles Symposium (IV), pages

724–731. IEEE.

Musabini, A., Novikov, I., Soula, S., Leonet, C., Wang, L.,

Benmokhtar, R., Burger, F., Boulay, T., and Perrotton,

X. (2024). Enhanced parking perception by multi-task

ﬁsheye cross-view transformers.

Pan, X., Tewari, A., Leimk

uhler, T., Liu, L., Meka, A., and

Theobalt, C. (2023). Drag your gan: Interactive point-

based manipulation on the generative image manifold.

In ACM SIGGRAPH 2023 Conference Proceedings.

Parallel Domain Plateform (n.d.). Parallel domain. Ac-

cessed: Febuary 2024.

Pham, T., Maghoumi, M., Jiang, W., Jujjavarapu, B. S. S.,

Sajjadi, M., Liu, X., Lin, H.-C., Chen, B.-J., Truong,

G., Fang, C., et al. (2024). Nvautonet: Fast and accu-

rate 360deg 3d visual perception for self driving. In

Proceedings of the IEEE/CVF Winter Conference on

Applications of Computer Vision, pages 7376–7385.

Regmi, K. and Borji, A. (2019). Cross-view image synthe-

sis using geometry-guided conditional gans. Com-

puter Vision and Image Understanding, 187:102788.

Reiher, L., Lampe, B., and Eckstein, L. (2020). A sim2real

deep learning approach for the transformation of im-

ages from multiple vehicle-mounted cameras to a se-

mantically segmented image in bird’s eye view. In

2020 IEEE 23rd International Conference on Intelli-

gent Transportation Systems (ITSC), pages 1–7.

Ren, B., Tang, H., and Sebe, N. (2021). Cascaded cross

mlp-mixer gans for cross-view image translation. In

British Machine Vision Conference.

Ren, B., Tang, H., Wang, Y., Li, X., Wang, W., and Sebe,

N. (2023). Pi-trans: Parallel-convmlp and implicit-

transformation based gan for cross-view image trans-

lation. In ICASSP 2023 - 2023 IEEE International

Conference on Acoustics, Speech and Signal Process-

ing (ICASSP), pages 1–5.

Samani, E., Tao, F., Dasari, H., Ding, S., and Banerjee,

A. (2023). F2bev: Bird’s eye view generation from

surround-view ﬁsheye camera images for automated

driving. 2023 IEEE/RSJ International Conference on

Intelligent Robots and Systems (IROS), pages 9367–

9374.

Smith, L. and Topin, N. (2018). Super-convergence: very

fast training of neural networks using large learning

rates. In Defense + Commercial Sensing.

Tan, M. and Le, Q. V. (2021). Efﬁcientnetv2: Smaller mod-

els and faster training.

Tang, H., Xu, D., Sebe, N., Wang, Y., Corso, J. J., and

Yan, Y. (2019). Multi-channel attention selection gan

with cascaded semantic guidance for cross-view im-

age translation. 2019 IEEE/CVF Conference on Com-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

162

puter Vision and Pattern Recognition (CVPR), pages

2412–2421.

Tang, H., Xu, D., Yan, Y., Torr, P. H. S., and Sebe, N.

(2020a). Local class-speciﬁc and global image-level

generative adversarial networks for semantic-guided

scene generation. In 2020 IEEE/CVF Conference

on Computer Vision and Pattern Recognition (CVPR),

pages 7867–7876.

Tang, W., Li, G., Bao, X., Nian, F., and Li, T. (2020b).

Mscgan: Multi-scale conditional generative adversar-

ial networks for person image generation. In 2020

Chinese Control And Decision Conference (CCDC),

pages 1440–1445.

TrueCar (n.d.). Truecar. Accessed: Febuary 2024.

Wang, L., Musabini, A., Leonet, C., Benmokhtar, R., Bre-

heret, A., Yedes, C., B

urger, F., Boulay, T., and Per-

rotton, X. (2023). Holistic parking slot detection with

polygon-shaped representations. In 2023 IEEE/RSJ

International Conference on Intelligent Robots and

Systems (IROS), pages 5797–5803. IEEE.

Wang, T.-C., Liu, M.-Y., Zhu, J.-Y., Tao, A., Kautz, J., and

Catanzaro, B. (2018). High-resolution image synthe-

sis and semantic manipulation with conditional gans.

In Proceedings of the IEEE Conference on Computer

Vision and Pattern Recognition.

Wang, Z., Simoncelli, E., and Bovik, A. (2003). Multiscale

structural similarity for image quality assessment. In

The Thrity-Seventh Asilomar Conference on Signals,

Systems & Computers, 2003, volume 2, pages 1398 –

1402 Vol.2.

Wu, S., Tang, H., Jing, X.-Y., Zhao, H., Qian, J., Sebe, N.,

and Yan, Y. (2023). Cross-view panorama image syn-

thesis. IEEE Transactions on Multimedia, 25:3546–

3559.

Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z.,

Huang, G., Li, H., Qiao, Y., Lu, L., et al. (2023).

Bevformer v2: Adapting modern image backbones to

bird’s-eye-view recognition via perspective supervi-

sion. In Proceedings of the IEEE/CVF Conference

on Computer Vision and Pattern Recognition, pages

17830–17839.

Yang, W., Li, Q., Liu, W., Yu, Y., Ma, Y., He, S., and Pan,

J. (2021). Projecting your view attentively: Monocu-

lar road scene layout estimation via cross-view trans-

formation. In 2021 IEEE/CVF Conference on Com-

puter Vision and Pattern Recognition (CVPR), pages

15531–15540.

Zhang, H., Xu, T., Li, H., Zhang, S., Wang, X., Huang,

X., and Metaxas, D. (2017). Stackgan: Text to photo-

realistic image synthesis with stacked generative ad-

versarial networks. In ICCV.

Zhang, L., Huang, J., Li, X., and Xiong, L. (2018). Vision-

based parking-slot detection: A dcnn-based approach

and a large-scale benchmark dataset. IEEE Transac-

tions on Image Processing, 27(11):5350–5364.

Zhou, B. and Krahenbuhl, P. (2022). Cross-view trans-

formers for real-time map-view semantic segmenta-

tion. In 2022 IEEE/CVF Conference on Computer Vi-

sion and Pattern Recognition (CVPR), pages 13750–

13759, Los Alamitos, CA, USA. IEEE Computer So-

ciety.

Zhu, J.-Y., Park, T., Isola, P., and Efros, A. A. (2017).

Unpaired image-to-image translation using cycle-

consistent adversarial networkss. In Computer Vision

(ICCV), 2017 IEEE International Conference on.

Zhu, X., Yin, Z., Shi, J., Li, H., and Lin, D. (2018). Genera-

tive adversarial frontal view to bird view synthesis. In

2018 International Conference on 3D Vision (3DV),

pages 454–463.

Zou, J., Xiao, J., Zhu, Z., Huang, J., Huang, G., Du, D., and

Wang, X. (2023). Hft: Lifting perspective represen-

tations via hybrid feature transformation for bev per-

ception. In 2023 IEEE International Conference on

Robotics and Automation (ICRA), pages 7046–7053.

BevGAN: Generative Fisheye Cross-View Transformers

163