Object Detection in Floor Plans for Automated VR Environment

Generation

Timoth

ee Fr

eville, Charles Hamesse, B

enoit Pairet and Rob Haelterman

Royal Military Academy, Rue Hobbema 8, Brussels, Belgium

Keywords:

Image Recognition, Floor Plans, Neural Networks, Synthetic Data.

Abstract:

The development of visually compelling Virtual Reality (VR) environments for serious games is a complex

task. Most environments are designed using game engines such as Unity or Unreal Engine and require hours

if not days of work. However, most important information of indoor environments can be represented by ﬂoor

plans. Those have been used in architecture for centuries as a fast and reliable way of depicting building

conﬁgurations. Therefore, the idea of easing the creation of VR ready environments using ﬂoor plans is of

great interest. In this paper we propose an automated framework to detect and classify objects in ﬂoor plans

using a neural network trained with a custom ﬂoor plan dataset generator. We evaluate our system on three

ﬂoor plans datasets: ROBIN (labelled), PFG (our own Procedural Floor plan Generation method) and 100

labelled samples from the CubiCasa Dataset

1 INTRODUCTION

Designing high-quality VR environments is a noto-

riously tedious but important task for the succesful

deployment of VR applications. Depending on the

target application, different levels of attention must

be brought to different aspects of the environment.

Aiming for maximum photorealism is not always the

best option, for example in the case of serious gaming

or procedure training applications where the place,

scale, behaviour and lifecycle of objects are the most

important. For example, in the case of ﬁreﬁght-

ing training, rooms must feature doors, windows and

ﬂammable props at relevant locations chosen by the

instructor (Haelterman et al., 2020).

Floor plans are an easily understandable and ed-

itable format. This representation also has the advan-

tage to be universal. Furthermore, with the democra-

tization of the use of software to design these plans,

architects and individuals are now able to create and

modify ﬂoor plans in a few minutes. Adding a VR

ready experience to this system is a natural prolonga-

tion of the technology. Furthermore, displaying inte-

rior layout using a VR headset can have much more

impact than a regular computer screen (Portman et al.,

2015) and (Davila Delgado et al., 2020).

In this work, we extend the framework proposed

in (Fr

eville et al., 2021). More speciﬁcally, we

present enhancements done on the object detection

component of the framework. In this case, the most

important aspect is the presence and placement of ob-

jects that will inﬂuence the progression of tactical

squad: doors, windows, room and building layouts.

Our contributions are the following:

• A ﬂoor plan dataset generator with various image

deterioration methods (noise, distortion, etc) and

ideal ground truth, to train a neural network with

robustness;

• A neural network trained on this dataset for a

more versatile detection and a thorough evalua-

tion on real-life ﬂoor plans;

• The integration of our method in a more generic

framework for converting ﬂoor plans into VR en-

vironments.

2 RELATED WORK

We review the state-of-the-art methods in ﬂoor plan

parsing, 3D environment generation for VR and com-

plete systems to generate VR environments from ﬂoor

plans.

2.1 Floor Plan Parsing

2.1.1 Classical Approaches

Traditional ﬂoor plan parsing techniques use con-

ventional image processing techniques to target ele-

480

Fréville, T., Hamesse, C., Pairet, B. and Haelterman, R.

Object Detection in Floor Plans for Automated VR Environment Generation.

DOI: 10.5220/0011629300003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages

480-486

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

ments that could point to the presence of walls, doors,

or other such things. (Mac

e et al., 2010) performs

line detection using image vectorization and a Hough

transform. If a particular graphical arrangement is

met, the lines are merged to form walls. The authors

suggest a similar technique to extract arcs and ﬁnd

door hypotheses. (de las Heras et al., 2011) proposes

an alternative approach that does not require image

vectorization and employs patch-based segmentation

with visual words. The method in (Ahmed et al.,

2011) is built to distinguish between thick, medium,

and thin lines in order to identify walls and eliminate

any components that are outside the convex hull of

the outer walls. In (Daniel Westberg, 2019), a thick

and a thin representation of the walls are extracted us-

ing noise removal techniques, erosions, and dilations.

Then, a hollow representation of walls with a constant

thickness is created by subtracting the thin represen-

tation from the thick one, which may then be used as

a reference for 3D modeling.

2.1.2 Deep Learning-Based Methods

To learn to predict room-boundary elements (e.g.

walls, windows, and doors) (Zeng et al., 2019) de-

scribes a multi-task neural network. In order to extract

features from the ﬂoor plan image, a shared encoder

is used, and one decoder is used for each task. Both

the encoder and the decoders’ architectures are based

on VGG (Simonyan and Zisserman, 2015). A Faster-

RCNN-based (Ren et al., 2015) object detector is en-

hanced in (Ziran and Marinai, 2018) to learn to antic-

ipate annotations in diverse ﬂoor plan datasets. The

same strategy is suggested in(Singh, 2019), however

this time it uses of a modiﬁed version of the YOLO

object detector (Redmon and Farhadi, 2017). Train-

ing or evaluation of these neural networks is done on

the datasets (Mathieu Delalandre, 2019) and (Chiran-

joy Chattopadhyay , 2019), which we also use in this

study.

2.2 3D Environment Generation

The game industry has been a major driver of ad-

vancement in the ﬁeld of 3D environment generation

during the past few years. 3D engines like Unity

(Haas, 2014) and Unreal Engine (Epic Games, 2019a)

are becoming easier to use, more adaptable, and more

potent. Environments can be created using 3D props

that have been manually or programmatically created

(and, if necessary, animated). The market has a huge

selection of 3D models, which facilitates the creation

of new surroundings. Blender (Foundation, 2002) is

also a good choice. It is open-source and has an active

community and plugins to execute tasks programmat-

ically from Python scripts. Since Blender does not

have as many interaction features or VR capabilities

like game engines, we opt to develop our algorithm

with game engines. More particularly we choose Un-

real Engine 4 because of its Virtual Reality integration

plugin that works perfectly with Steam VR (Valve,

2003), ensuring the versatility of our model on var-

ious VR platforms such as Oculus (Meta, 2012) or

Varjo (Varjo, 2016). Furthermore, the access to the

Unreal Market (Epic Games, 2019b) is an important

feature that allows us to ease the process of our im-

plementation by getting plugins and assets from the

community.

2.3 Floor Plan to VR Environment

Frameworks

There exists many frameworks that achieve the tasks

of ﬂoor plan parsing and 3D environment genera-

tion. 3DPlanNet uses heuristic rules to retrieves walls

and Tensorﬂow Object detection API for windows

and doors (Park and Kim, 2021). This study (Dodge

et al., 2017) uses a fully connected neural network

and uses OCR (Optical Character Recognition) to es-

timate the size of the rooms. The framework proposed

in (Fr

eville et al., 2021) combines traditional com-

puter vision and deep-learning techniques to detect

room boundary features (walls, doors and windows,

interior objects). They implement ad-hoc map gener-

ation scripts for Unreal Engine to turn ﬂoor plans into

VR-ready environments. In our case, we use a neural

network-based solution for object detection using a

Yolo instance (Redmon and Farhadi, ). To ensure suf-

ﬁcient training data, we design a ﬂoor plan generator

with random image perturbations. Doing so allows us

to feed YoloV5 with virtually unlimited training data,

allowing our implementation to be more robust.

3 METHOD

In this section, we present our ﬂoor plan parser and

our custom dataset generator. The architecture of the

parser system, depicted in Figure 3 has two main two

parts :

1. The furniture recognition : which uses the

YOLOv5 Framework (Deep Learning methods)

2. The wall recognition that uses OpenCV libraries

(Traditional image processing tool)

Object Detection in Floor Plans for Automated VR Environment Generation

481

3.1 Walls Detection

Our method is based on (Daniel Westberg, 2019),

which mainly uses traditional computer vision meth-

ods and is implemented using OpenCV. The steps of

the procedure to go from a grayscale ﬂoor plan image

to a list of wall coordinates are the following:

1. Conversion to a binary image using binary inverse

thresholding.

2. Noise removal method with opening (erosion then

dilation) to remove the thin details (e.g. furniture).

At this point, we have an image representing a

thick version of the walls.

3. Distance transform: Each pixel that is a part of a

wall has its value changed to reﬂect the separation

between it and the closest black pixel, or the wall’s

edge.

4. This distance image is converted to a binary image

once again using binary thresholding. As a result,

the walls appear narrow in the resulting image.

5. A representation of hollow walls with a constant

thickness is produced by subtracting thin walls

(Step 4) from thick walls (Step 2) in the process.

3.2 Object Detection

The most recent edition of YOLO is the v5 (Red-

mon and Farhadi, ). The backbone is composed of a

Cross-stage partial networks CSP (Wang et al., 2019)

are employed in YOLOv5 to extract valuable fea-

tures. For feature pyramid retrieval, YOLOv5 lever-

ages PANet (Liu et al., 2018). Then, YOLOv5 uses

the sigmoid activation function in the ﬁnal detection

layer and the leaky ReLU activation function in the

middle/hidden layers. For naturally occurring photos,

YoloV5 is one of the most rapid and precise object

detection method.

Figure 1: YoloV5 on inference.

YoloV5 has shown its effectiveness to recognize

2D features on hand-drawn images (Reddy and Pan-

icker, 2021). Therefore this neural network frame-

work has been chosen for the furniture recognition

part. The training was done using a multi-GPU setup

using GTX Tesla cards, a batch size of 16, learning

rate of 0.001, a momentum of 0.937 and 500 epochs.

Using YOLOv5 allows us to beneﬁt from the exten-

sive inference and test metrics implemented by de-

fault in the framework.

3.3 3D Environment Generation

The 3D generation is achieved using Unreal Engine 4.

Therefore we used a plugin found on the marketplace

to import the meshes in runtime (Virtual Bird, 2018).

Furthermore we create our own plugin to transform

YOLOv5 text ﬁle output into readable object by Un-

real Engine.

3.4 Training Data Generator

To train a deep neural network we need numerous

labeled images. Nowadays many datasets are avail-

able online for common object detection (e.g: air-

planes, cars, boat, animals, etc) on various websites

such as Kaggle, Robotﬂow or Github. This consid-

erably eases the neural network learning process as

the quality of the neural network highly depends on

the data feed. Unfortunately, speciﬁc object detection

requires speciﬁc datasets. Therefore, a vast quantity

of labeled ﬂoor plans are required for our application.

This leads to two issues. First, this sort of dataset is

relatively hard to ﬁnd. Second, ﬂoor plan data found

in one source are most likely to use the same sym-

bols for furniture which increases the probability of

over-ﬁtting and does not help our model to be robust

during inference. Our solution was to program a ﬂoor

plan generator that can create thousands of labelled

data using different layouts.

Figure 2: Example of ﬂoor plan generated.

Creating an accurate and realistic ﬂoor plan im-

age by using a generator can quickly become a tedious

task if we do not restrain the scope on the most impor-

tant aspect: interior object recognition. The aim here

is to merge the window and door recognition with the

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

482

furniture recognition. Therefore, creating an image

with a convincing room conﬁgurations and usable fur-

niture arrangement does not enter into our considera-

tion. The neural network only needs to learn object

shapes and basic placement rules (e.g. doors and win-

dows are mostly linked with walls). Using the known

rules and random functions we can develop a program

to generate thousands of labeled ﬂoor plans with dif-

ferent ﬂoor layouts and symbols.

3.5 Complete System

Figure 3 represent the ﬁnal architecture of our

method. Using multiple python scripts in parallel and

Docker we can easily generate a 3D representation of

the world using Unreal Engine 4 as our main virtual

engine.

Figure 3: Final Architecture.

4 EVALUATION

The military end-users would have different level of

detail of ﬂoorplans in their possession.

• Some of them would be accessible directly from

the architect plan achieved using a software.

Which means almost zeros noise and high reso-

lution plans. We are testing this using a synthetic

dataset named ROBIN.

• The second option would be old plans scanned,

which obviously features more noise and defaults

induced by the age of the paper itself or the draw-

ing done by the architect. We would be testing

it using Cubicasa datasets which feature scans of

real-world ﬂoor plans.

• The ﬁnal option would be to draw by hand directly

on paper the ﬂoorplan if no other option is avail-

able. This features becomes quite interesting, al-

lowing the end-user to draw virtually hundred of

plans for tactical training. For this we will test on

drawn ﬂoor plan made by our self.

We evaluate the performance of our methods on a

machine with an Intel i7 8-core CPU and an Nvidia

RTX 2080Ti GPU. The computational time required

to carry out our method is extremely brief: only a few

seconds are needed to run the detection procedure,

which includes the neural network inference. The

use of Docker allows us to use pre-conﬁgured

environments with libraries and tools to run codes.

Using this technique we greatly enhance the number

of VR environments generated. The YOLO instance

is launched to detect every image present on a folder.

Since the implementation is splitting the generation

of walls and the furniture detection, it is much more

efﬁcient and reliable to only evaluate our model on

the furniture part. Firstly, it is a tedious task to ﬁnd a

ﬂoor plan dataset that has walls labelled. Therefore

the inference on furniture would be our only metrics.

Our model only involves the position and the size of

the objects found, not rotation. However, in our use

case the rotation can be easily guessed as objects such

as sinks, couches or toilets are in most apartments

turned towards the center of the room. Furthermore,

objects as tables or plants do not need rotation

information to be accurate. The evaluation part has

been split in two parts: ﬁrst our detection system is

evaluated with the ROBIN synthetic dataset. In a

second time it will be evaluated with a real-world

dataset. Those datasets will be scans of buildings

available from real estate companies. Three different

metrics have been used for the evaluation part:

mAP: Mean Average Precision : The mAP cal-

culates a score by comparing the detected box to the

ground-truth bounding box. The model’s detections

are more precise the higher the score.

Precision: Precision is a model’s capacity to recog-

nize only the pertinent objects. A model with a preci-

sion of 1.0 produces no false positives. However, even

if there are bounding boxes that should be identiﬁed

but aren’t, the value will still be 1.0.

Precision = T P/(T P + FP) (1)

Recall: Recall is a model’s capacity to locate every

ground truth bounding box. A model with a recall of

1.0 creates no false negatives, or undetected bounding

boxes that should be identiﬁed. The recall will still be

1.0, even if there is a ”overdetection” and the incorrect

bounding box is discovered.

Recall = T P/(T P + FN) (2)

4.1 Synthetic Data (ROBIN)

Many research papers use the ROBIN dataset to

validate their framework: (Sharma et al., 2018) uses

Object Detection in Floor Plans for Automated VR Environment Generation

483

it to test the performance of their neural network

to match ﬂoorplans with hand-drawn sketch plans.

(Goyal et al., 2019) uses the dataset to generate more

ﬂoor plans. Better results are expected with this type

of data since it is noise- and distortion-free. The la-

belling of this dataset has been done using Robotﬂow

and can be accessed in (Timoth

ee Fr

eville, 2022) for

further investigation. Therefore, it is a perfect case to

test our algorithms on a synthetic dataset that was not

in the training set. If the inference test is satisfactory

we could switch to a real world dataset.

Table 1: Results on ROBIN datasets.

Object mAP Precision Recall

Doors 0.774 0.713 0.771

Table 0.361 0.695 0.082

Sink 0.254 0.546 0.327

Couch 0.496 0.643 0.259

Bed 0.280 0.317 0.699

Toilet 0.738 0.684 0.759

Tub 0.743 0.650 0.836

All 0.536 0.606 0.533

Figure 4: Sample of the ROBIN dataset.

The ROBIN dataset does not feature window ob-

jects. As we can see the model confuses objects. This

can be explained by the symbol system used by the

ROBIN dataset that differs signiﬁcantly from the ones

present in the dataset generated to train YOLOv5 (e.g.

tables are represented by a square with a cross which,

model has never been trained with such representa-

tion). Furthermore, sinks and beds differ a lot com-

pared to the training dataset used Fig5. However, the

model detects almost every object on plans but the

recognition can deﬁnitely be enhanced.

Figure 5: Objects that are hardly recognized by our frame-

work.

4.2 Real-World Data (CubiCasa)

The CubiCasa dataset does not feature objects of

couch, bed and toilet classes. In particular, the

CubiCasa5K version is a large-scale ﬂoor plan image

collection with 5000 samples with objects that have

been classiﬁed into more than 80 categories. By

employing polygons to divide the various objects, the

dataset annotations are more precise than the usual

ones using bounding boxes. This dataset suits our

need of diverse symbols and ﬂoor plan layouts in

general. Most of the images feature distortion and

noise as most of the images are pictures or scans

taken from the real plans.

Table 2: Results on CubiCasa datasets.

Object mAP Precision Recall

Window 0.353 0.585 0.162

Doors 0.461 0.514 0.408

Sink 0.107 0.285 0.027

Toilet 0.074 0.153 0.016

Tub 0.016 0.047 0.021

All 0.202 0.316 0.126

Figure 6: Sample of the Cubicasa dataset.

As expected, the model gets more frequently lost

against a real-world dataset. Most of the objects are

confused with the background. The real-world dataset

features pen-drawing, distortion, colors for certain

plan and an overlapping of symbol that confuses the

recognition.

4.3 Hand-Drawn Dataset

This hand-drawn dataset has been created specially

for this study. This is a feature specially interesting

for the military forces, as they can use it draw the ﬂoor

worthy of interest and quickly train for intervention.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

484

This dataset is composed of seven plans that fea-

ture objects that resemble the ones used by the ﬂoor

plan generator. Five of them are archived using a ruler

for a better precision and two with free-hand drawing

to challenge the neural network.

Table 3: Results on hand drawing datasets

Object mAP Precision Recall

Window 0 0 0

Doors 0.840 0.913 0.778

table 0.900 1 0.800

Sink 0.687 1 0.375

couch 0.812 1 0.625

bed 0.750 1 0.500

Toilet 0.642 0.750 0.429

Tub 0.723 0.600 0.750

All 0.669 0.782 0.532

Figure 7: Sample of the Hand-drawn dataset.

Our method yields good results on the hand-drawn

dataset. This indicates that the patterns learnt by the

neural network during the training is robust enough to

understand a hand-drawn version of a ﬂoor plan. This

results are really interesting to validate our model in

a near future. However, a dataset of seven plans is

not enough to draw conclusion on the validity of our

model on other hand-drawn ﬂoor plans.

5 CONCLUSION

In this paper we demonstrated an automated frame-

work to detect and classify objects in ﬂoor plans us-

ing a neural network trained with a custom ﬂoor plan

dataset generator. Our custom ﬂoor plan dataset gen-

erator allows our neural network to be useful on dif-

ferent plan representations. As presented in the sec-

tion 4.1 and 4.2 the model performs better on syn-

thetic dataset than the real world dataset. The noise

and blur induced on realistic ﬂoorplan confuses the

YOLOv5 instance that has only been fed with syn-

thetic procedural data. A more robust model can

be build using a more versatile layout and symbol-

ism on ﬂoorplan to generate and including real world

datasets in the training process. More variation of the

same images with blur and distortion can be added to

achieve better results as well.

REFERENCES

Ahmed, S., Liwicki, M., Weber, M., and Dengel, A. (2011).

Improved automatic analysis of architectural ﬂoor

plans. In 2011 International Conference on Document

Analysis and Recognition, pages 864–869.

Chiranjoy Chattopadhyay (2019). Repository of building

plans (robin).

Daniel Westberg (2019). Floor plans to blender 3d. https:

//github.com/grebtsew/FloorplanToBlender3d/.

Davila Delgado, J. M., Oyedele, L., Demian, P., and Beach,

T. (2020). A research agenda for augmented and vir-

tual reality in architecture, engineering and construc-

tion. Advanced Engineering Informatics, 45:101122.

de las Heras, L.-P., Mas, J., S

anchez, G., and Valveny, E.

(2011). Wall patch-based segmentation in architec-

tural ﬂoorplans. In 2011 International Conference

on Document Analysis and Recognition, pages 1270–

1274.

Dodge, S., Xu, J., and Stenger, B. (2017). Parsing ﬂoor

plan images. In 2017 Fifteenth IAPR International

Conference on Machine Vision Applications (MVA),

pages 358–361.

Epic Games (2019a). Unreal engine. https://www.

unrealengine.com.

Epic Games (2019b). Unreal market.

Foundation, T. B. (2002). Blender. https://www.blender.

org/.

eville, T., Hamesse, C., Pairet, B., Lahouli, R., and Hael-

terman, R. (2021). From ﬂoor plans to virtual reality.

In 2021 IEEE International Conference on Artiﬁcial

Intelligence and Virtual Reality (AIVR), pages 129–

133.

Goyal, S., Mistry, V., Chattopadhyay, C., and Bhatnagar, G.

(2019). Bridge: Building plan repository for image

description generation, and evaluation.

Haas, J. K. (2014). A history of the unity game engine.

Worcester Polytechnic Institute.

Haelterman, R., Bellemans, M., Lammens, D., Sloover, J.,

De Vleeschauwer, T., Schoofs, E., Jordens, W., Steen-

huyse, B., Mangelschots, J., Selleri, S., Hamesse, C.,

and Fr

eville, T. (2020). Training ﬁreﬁghters in virtual

reality.

Liu, S., Qi, L., Qin, H., Shi, J., and Jia, J. (2018). Path ag-

gregation network for instance segmentation. CoRR,

abs/1803.01534.

Mac

e, S., Locteau, H., Valveny, E., and Tabbone, S. (2010).

A system to detect rooms in architectural ﬂoor plan

images. pages 167–174.

Mathieu Delalandre (2019). Systems evaluation synthetic

documents (sesyd). http://mathieu.delalandre.free.fr/

projects/sesyd/.

Object Detection in Floor Plans for Automated VR Environment Generation

485

Meta (2012). oculus. https://www.oculus.com/rift-s/

?locale=fr FR.

Park, S. and Kim, H. (2021). 3dplannet: Generating 3d

models from 2d ﬂoor plan images using ensemble

methods. Electronics, 10:2729.

Portman, M., Natapov, A., and Fisher-Gewirtzman, D.

(2015). To go where no man has gone before: Virtual

reality in architecture, landscape architecture and en-

vironmental planning. Computers, Environment and

Urban Systems, 54:376–384.

Reddy, R. R. and Panicker, M. R. (2021). Hand-drawn elec-

trical circuit recognition using object detection and

node recognition. CoRR, abs/2106.11559.

Redmon, J. and Farhadi, A. Yolov3: An incremental im-

provement.

Redmon, J. and Farhadi, A. (2017). Yolo9000: Better,

faster, stronger. pages 6517–6525.

Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster

r-cnn: Towards real-time object detection with region

proposal networks. In Proceedings of the 28th Inter-

national Conference on Neural Information Process-

ing Systems - Volume 1, NIPS’15, page 91–99, Cam-

bridge, MA, USA. MIT Press.

Sharma, D., Gupta, N., Chattopadhyay, C., and Mehta, S.

(2018). Rexplore: A sketch based interactive explorer

for real estates using building ﬂoor plan images. In

2018 IEEE International Symposium on Multimedia

(ISM), pages 61–64.

Simonyan, K. and Zisserman, A. (2015). Very deep con-

volutional networks for large-scale image recognition.

In International Conference on Learning Representa-

tions.

Singh, D. (2019). Object detection in ﬂoor plan images.

Timoth

ee Fr

eville (2022). Datasets for evaluation.

https://github.com/Tim-HW/Object-detection-

in-ﬂoor-plans-for-automated-VR-environment-

generation-datasets.

Valve (2003). Steam vr. https://store.steampowered.com/

app/250820/SteamVR/.

Varjo (2016). Varjo. https://varjo.com/.

Virtual Bird (2018). Import runtime ue4. https://www.

virtualbird.de/ue4Doku/RealTimeImport/.

Wang, C., Liao, H. M., Yeh, I., Wu, Y., Chen, P., and Hsieh,

J. (2019). Cspnet: A new backbone that can enhance

learning capability of CNN. CoRR, abs/1911.11929.

Zeng, Z., Li, X., Yu, Y., and Fu, C. (2019). Deep ﬂoor plan

recognition using a multi-task network with room-

boundary-guided attention. In 2019 IEEE/CVF In-

ternational Conference on Computer Vision (ICCV),

pages 9095–9103, Los Alamitos, CA, USA. IEEE

Computer Society.

Ziran, Z. and Marinai, S. (2018). Object detection in ﬂoor

plan images. In Pancioni, L., Schwenker, F., and

Trentin, E., editors, Artiﬁcial Neural Networks in Pat-

tern Recognition. Springer International Publishing.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

486