=Paper=
{{Paper
|id=Vol-3869/p07
|storemode=property
|title=A Real-Time Machine Learning Based Solution for Privacy
Enforcement in Video Recordings and Live Streaming
|pdfUrl=https://ceur-ws.org/Vol-3869/p07.pdf
|volume=Vol-3869
|authors=Pietro Manganelli Conforti,Matteo Emanuele,Lorenzo Mandelli
|dblpUrl=https://dblp.org/rec/conf/icyrime/ConfortiEM24
}}
==A Real-Time Machine Learning Based Solution for Privacy
Enforcement in Video Recordings and Live Streaming==
https://ceur-ws.org/Vol-3869/p07.pdf
A Real-Time Machine Learning Based Solution for Privacy
Enforcement in Video Recordings and Live Streaming
Pietro Manganelli Conforti1 , Matteo Emanuele1 and Lorenzo Mandelli1
1
Department of Computer, Control and Management Engineering, Sapienza University of Rome, Italy
Abstract
These past years the world had to deal with a whole new situation brought by Covid-19. Everyone’s routine changed
and we started passing way more time than before on virtual meeting, virtual chats and similar. With this, many privacy
problems arised from all the video data generated by a single user. Google and Zoom introduced the possibility to blur out
the background while using a front face camera, but this did not solve many privacy concerns ranging from showing people
in videos without their permission, to the leaking of sensible data and information from videos uploaded online. We propose
a solution build over the use of computer vision techniques like image segmentation and classification for context recognition
for a privacy enforcement solution capable of fitting the user’s personal need, blurring out selectively specific objects from a
video based on the user’s preferences for each room in which they are.
Keywords
Image segmentation, Context Recognition, Detectron2, Privacy enforcement, Covid-19, Alexnet, Transfer learning,
1. Introduction to show, based on the recognition of the environment
framed, in order to blur out objects relatively to both the
In the past years there has been a solid shift for the en- user needs and the context in which they are.
tire world population towards a more active presence
online. Covid-19 has further pushed many activities to
be faced digitally. Virtual meeting application like Zoom,
had 10 million daily meeting participants in December 2. Related works
2019, but by April 2020, that number increased to reach
up to 300 million [1]. It is estimated that in 2024 only With the advancement of technology, people have been
25% of the business meetings will take place in person sharing a continuously growing amount of personal data
[2]. Studies started during 2020 have demonstrated that online. Additionally to life-logging devices[9], social
nowadays people spend on average way more time in vir- medias have recently stepped in, ending up quickly dom-
tual meetings than before [3], leading to many concerns inating the landscape of mass produced data with "visual
for the single person. Users have started experiencing data"(i.e. images and video). For instance, in 2020, the
stress related to not being competent in the use of the first year of pandemic, users have generated and shared
technology, but most importantly to "Zoom fatigue" due via Facebook a total of 10.5million videos[10].
to it being always "on" [4]. Many privacy related issues This impactful amount of data brought to the attention
have been crippling the user experience ever since, such of experts and users to many privacy related issues; stud-
as exposing private and personal spaces on camera, un- ies started identifying and observing how easily privacy
intentionally framing a person who did not give consent could be violated just from unintentionally sharing per-
to be on video or sharing sensible information leaked sonal data contained inside images and videos, and sub-
from careless online posting. Many solutions have been sequently started proposing privacy models to formally
promptly developed to prevent such things from hap- approach and tackle said scenarios [11]. The scientific
pening, providing virtual meeting room services with world went quickly from defining sub-fields like Privacy-
safeguard-privacy functionalities like blurring the back- Preserving Machine Learning(PPML) [12], to adopting
ground and virtual backgrounds[5, 6, 7, 8]. We present deep learning models for image disguising [13], Context
in this paper a novel computer vision based approach for Recognition[14] as well as Image-Based Localization[15]
privacy enforcement in video data, capable of filter out and again computer vision based framework[16] as novel
from a video a list of objects that a user does not want solutions for privacy preserving of first person vision
image sequences, placing computer vision, Artificial In-
telligence and data driven approaches as state of the
ICYRIME 2024: 9th International Conference of Yearly Reports on
Informatics, Mathematics, and Engineering. Catania, July 29-August art techniques for preserving privacy on line. Among
1, 2024 the many available solutions for privacy preservation
$ mandelli@diag.uniroma1.it (L. Mandelli) and safeguarding, it is currently missing one which al-
© 2024 Copyright for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0). lows single users to selectively censor objects from visual
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org)
53
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
Pietro Manganelli Conforti et al. CEUR Workshop Proceedings 53–59
Figure 1: The pipeline of our system.
data depending on their personal needs and preferences. very powerful instance segmentation network published
The propose work aims therefore to provide experience- by Facebook in 2019 [27], is used to identify user’s con-
lacking users with an intuitive, easy to use tool for pri- text specific, privacy related data, within video frames,
vacy enforcement in video data based on computer vision with no ambiguity; combining the output masks pro-
techniques. duced by Detectron 2 with all the information retrieved
before, a particular region of the frame is identified and
filtered with a Gaussian transform. Similar or identical
contexts disambiguation is possible to be tackled and
3. Implementation solved with the support of RFID technology: with the
introduction of a beacon that send a constant signal, it
Sensible users’ data is crucial to be kept private. The
is possible to recognize and distinguish two apparently-
proposed software tracks such information by means of
same looking environment. Such discriminatory action
various modules which pipeline is shown in fig.1. Mem-
is essential, yet simple to be applied since it is integrable
ory buffers are used in between modules to guarantee
in any environment with low effort or invasiveness. A
flexibility towards input videos of any 𝑎𝑠𝑝𝑒𝑐𝑡, 𝑟𝑎𝑡𝑖𝑜 and
similar RFID-based solution for context recognition was
𝑓 𝑝𝑠, as well as to stabilize the output overcoming the
already presented by another research [14] conducted
common flickering experienced in these kind of appli-
some years ago. Finally the desired effect is obtained by
cations. Thanks to such buffers it is possible to store
processing and collecting all the frames of the video and
past frames and reuse those to statistically smooth of the
setting the right frame velocity.
final output; past frames are reused according to level of
confidence the class recognition module predicted with.
Input videos can be directly uploaded to the system or
streamed from cameras(i.e. webcams). 3.1. Dataset
The proposed solution separates the overarching learning Distinct datasets have been used for the two different
problem in two sub problems, namely context recogni- learning tasks respectively, image segmentation and con-
tion and image segmentation; this approach guarantees text recognition. The choice of using the Detectron2
robustness through modularity and simplifies the overall network for the the image segmentation tasks leaves lit-
functioning of the software. tle to no choice but using the 2017 version for the COCO
Recognizing the users, their emotional state [17, 18, 19], dataset [28] which has been demonstrated to be perfor-
his attentive state [20, 21, 22], and the context surroundings[23,
mative with such dataset.
24, 25] allows to selectively obscure specific elements COCO is a dataset composed of two groups of elements:
based on preferences the same user expressed at the reg- images and annotations. Images contain a vast variety
istration time; such data is stored into a database for later of objects, for a total of 80 different category of elements.
inference. Context recognition has been tackled with a The network was capable to recognize them all, and even
neural network inspired by Alexnet [26], a famous Deep apparently odd objects were left untouched and not re-
Convolutional Neural Network, designed for image clas- moved. Together with the set of images, COCO is com-
sification. posed of a set of so-called "annotations" that contain
Together with an RFID application [14] Detectron 2, a information related to the position of the object masks,
54
Pietro Manganelli Conforti et al. CEUR Workshop Proceedings 53–59
their bounding box and their location on the image ref- image to classify. For this task Alexnet has been fine
erence frame. tuned on our reduced dataset by means of transfer learn-
For what concern the context recognition part, a slightly ing.
modified version of a dataset available on Kaggle[29] has The structure of the network is shown in figure 2.
been used; such dataset is composed of 5 different classes
symbolizing five different kind of rooms, two of which Here we report the performance with which we eval-
has been merged together, namely living room and din- uated our model. We will give particular importance to
ing room. Each element is originally an RGB picture of both the accuracy and F1-score of each class.
a fixed size of 224x224x3 which has been resized to be
Classes precision recall F1-score overall Accuracy
227x227x3, to better fit through AlexNet (?).
Bathroom 0.84 0.90 0.87
As part of the training & testing process a defined set of LivingRoom 0.92 0.79 0.85
image processing techniques have been organized into a Bedroom 0.69 0.76 0.76
pipeline. Such transformation pipeline has been imple- Kitchen 0.75 0.76 0.76
mented using Albumentation library[30], an easy-to-use 0.83
and intuitive library for image processing; it consists of: Table 1
ShiftScaleRotate, for shifting or rotating images; RGB- Performance of our classification task
shift for randomly altering RGB channels’ values; Ran-
domBrightnessContrast for randomly changing iamges’ The performance of the network are reported in the
brightness and contrast; MultiplicativeNoise for randomly table 1. As we can see we are capable of obtaining high F1-
adding noise; Normalize, for normalizing data; HueSatu- score values for each class and an overall accuracy above
rationValue, for randomly changing images’ saturation. 80%, making the results satisfying for our standards. We
can consider the macro F1 score as a general metric of
evaluation, defined as:
3.2. Image Segmentation Network 1 ∑︁
𝑁
𝑚𝑎𝑐𝑟𝑜 − 𝐹 1 = 𝐹 1𝑖 = 0.81
Detectron2 is Facebook AI Research’s library [27] that 𝑁 𝑖=1
provides state-of-the-art detection and segmentation al-
Which is simply an overall F1 score among all classes. In
gorithms. It is the successor of Detectron [31] which is
our case, we can see a macro F1 score of 0.81.
in turn based on the Maskrcnn-benchmark model [32].
It supports a great number of computer vision research
It is also possible to take vision of the confusion matrix
projects thanks to its flexibility, output capabilities and
in figure 3, showing how the different samples from the
available documentation.
test set were classified during the test phase.
Among the available Detectron2’ architectures mask-
rcnn-fpn has been chosen. Such architecture is mainly
built from three modules: a Backbone Network, a Region
Proposal Network and a Box Head. 3.4. Output stabilization techniques
The 𝐵𝑎𝑐𝑘𝑏𝑜𝑛𝑒 𝑁 𝑒𝑡𝑤𝑜𝑟𝑘, whose role is to extract multi- Videos in daily life scenarios, likely happen to contain
scale feature maps with different receptive fields starting temporary blank frames, as well as artifacts, due to user
from the input image, is based on the Feature Pyramid or scene related conditions. A context recognition net-
Network [33] technique. In this way areas of interest work will therefore generate a classification label that
from different points of view are identified and passed will be trivially assigned, leading to an instability prob-
to both the two next modules. The 𝑅𝑒𝑔𝑖𝑜𝑛 𝑃 𝑟𝑜𝑝𝑜𝑠𝑎𝑙 lem. We dealt with this problem through the introduction
𝑁 𝑒𝑡𝑤𝑜𝑟𝑘 detects object regions (which are the so called of a "memory buffer", that is capable of stabilizing statis-
“proposal boxes” ) based on multi-scale features, which tically the result.
together with the feature maps serve as input for 𝑅𝑜𝑖
(Region of interest) 𝐻𝑒𝑎𝑑. This last module warps fea- This is achieved by endowing the system with two
ture maps using proposal boxes into multiple fixed-size buffers, one for the context recognition network that
features, and retrieves the fine-tuned box locations and stores the predicted context classes, and one used to track
classification results via fully-connected layers. the instance segmentation network output and to store
the classes predicted with enough accuracy. Storage of
past data allows to create a time relation between suc-
3.3. Context Recognition Network cessive frames, thus enforcing the output of each net-
work and stabilizing the final one. This method allows
The context recognition task belongs to a classification to correlate information inside the video with the least
problem, where each frame of the video is treated as an expenditure of resources. There are two buffers.
55
Pietro Manganelli Conforti et al. CEUR Workshop Proceedings 53–59
Figure 2: Alexnet architecture. All rights reserved to the owner of the picture[34]
to change the output and in this short period of time is
wrongly classified with the previous stable label. This is
strongly compensated by the stability provided and the
delay is short enough to be difficult to notice.
3.4.2. Instance segmentation memory buffer
The other output stabilization technique is the instance
segmentation memory buffer dedicated to the Detectron
output. The rationale here is regarding the threshold
used by the model to decide if an element belongs or not
to a certain class. Being a privacy concern to conceal
as much as possible the sensible information inside the
frames, false positives in exchange of a higher number
of true positives are preferable. Therefore, two kinds
of thresholds are considered: the basic one and the op-
timal one. The first one is lower than the second one
Figure 3: The confusion matrix generated with the use of the
scikit library [35] and is the minimal value considered acceptable to take
into account the output of the network. If the output
confidence regarding a specific instance inside the frame
falls below this value is considered too unclear and it is
3.4.1. Context memory buffer not counted. The second threshold instead represents
The first buffer visible in the top of fig.1 is the one ded- the optimal value of confidence used by the system in
icated to the output of Alexnet. The assumption taken order to properly recognize an element with enough ac-
by this approach is that the context changes don’t take curacy. This information is used in order to track the
place suddenly but instead are related to a smooth trend. last elements appeared to the network, appending them
For instance, if the frame 𝑛 recognizes a specific context, inside the buffer.
there is a high probability that also the frame 𝑛 + 1 will If the networks find an instance of a class inside the buffer
carry out similar information and represents the same even with a lower confidence value in respect to the op-
context. Thus, doing an average among the past frames timal threshold it is still considered acceptable, therefore
increases the overall accuracy by smoothing the output processed and eventually concealed. In this way it gets
trend. easier for the network to work with moving objects be-
The length of the buffer is set dynamically in relation to cause this method allows it to trace them even in the case
the 𝑓 𝑝𝑠 value retrieved from the video and the informa- of uncertainty due to movement.
tion obtained by the frames from the last half of second The buffer length is dynamically related to the 𝑓 𝑝𝑠 value
gets stored within. and stores information regarding the set of frames con-
The trade-off of this method is a small delay from the cerning the last three seconds. If an element is recog-
context recognition module because the output context nized by the network after this time interval in order to
label has to stabilize for at least half of the buffer 𝑙𝑒𝑛𝑔𝑡ℎ be evaluated it needs to overcome the optimal confidence
threshold again.
56
Pietro Manganelli Conforti et al. CEUR Workshop Proceedings 53–59
Figure 4: a: kitchen b: Bathroom c: Bedroom
The trade-off of this system, as mentioned above, is a mentation network recognize as a WC also the
higher frequency of false positives, which can be mis- bidet given the similarity in their structure.
leading for the final result and inversely proportional
in number to the two threshold values. Overall, the ac- • 𝐵𝑒𝑑𝑟𝑜𝑜𝑚 (fig. 4.c), In this scenario there is a
curacy following this approach improved by a discrete bowl placed in a flat surface behind the bed, and
percentage, mainly in the more dynamic scenarios. the user wants to blur it out. This scenario can
be potentially challenging since the portion of
the room we are framing is very restricted, and
the only object that can be considered a strong
feature is the bed.
4. Results
The results of the system are evaluated accordingly on The results are shown in the following table. Here we
how many times the full procedure works consistently to have 5 different values of evaluation for each test:
specific information given as input related to a specific
test video. Knowing in advance those information which • accuracy of context recognition network respect the
are the settings inside a set of test videos as well as the list total number of frames(C.R.), indicating the per-
of elements inside of them, we can measure the overall centage of success for the context recognition
accuracy of the system. network applied to the frames of the video. This
For instance, if a video displays a specific context with value does not show the improvements brought
a certain number of known elements inside of it we can by the memory buffer.
control how many times those elements are found by
the two networks and by the two memory buffers in the • accuracy of context recognition network + memory
output trend. buffer respect the total number of frames (C.R. ∖𝑤
In order to achieve this result an evaluation procedure B.), indicating the percentage of success for the
was implemented that given an input video follows sim- context recognition network combined with the
ilar steps that the system does but keeps into account memory buffer for the context recognition task.
the number of times the output given starting from each
input frame is correct in respect to the total number of • Accuracy of the instance segmentation network in
them. The test video used are three, which are providing finding the objects of interest in a frame respect
the following scenarios: the total number of frame(I.S.). This indicates the
• 𝑘𝑖𝑡𝑐ℎ𝑒𝑛 (fig. 4.a), where we want to blur out a percentage of success for the instance segmen-
bowl from a table given that the system recognize tation task respect the objects of interest for the
the context. The instance segmentation network user. This value does not show the improvements
can identify various objects as a table, a oven and brought by the memory buffer.
bottles. Due to the user preferences, here the sta-
tionary objects we want to blur out from all the • Accuracy of the instance segmentation network +
video frames is simply one, a black bowl. memory buffer in finding the objects of interest in a
frame respect the total number of frame(I.S. ∖𝑤 B.).
• 𝑏𝑎𝑡ℎ𝑟𝑜𝑜𝑚 (fig. 4.b), where the user want to blur this indicates the percentage of success for the
out the WC. In such scenario, the instance seg- instance segmentation task respect the objects of
57
Pietro Manganelli Conforti et al. CEUR Workshop Proceedings 53–59
interest for the user. This value does not show References
the improvements brought by the memory buffer.
[1] B. Evans, The zoom revolution: 10 eye-popping
• overall accuracy of the whole system. This indi- stats from tech’s new superstar, 2020.
cates, as the name states, the overall accuracy [2] W. Standaert, S. Muylle, A. Basu, How shall we
of the whole pipeline. This accuracy is given as meet? understanding the importance of meeting
a combined accuracy from the accuracy of the mode capabilities for different meeting objectives,
two tasks, obtained as the product between the Information Management (2021). doi:https://
accuracy of the instance segmentation consider- doi.org/10.1016/j.im.2020.103393.
ing also the buffer and the context recognition [3] D. Chew, M. Azizi, The state of video conferencing
considering also the buffer. 2022, 2022.
[4] K. A. Karl, J. V. Peluchette, N. Aghakhani, Virtual
work meetings during the covid-19 pandemic: The
In table 2 tests’ results are reported, confirming that mem- good, bad, and ugly, Small Group Research 0 (2021).
ory buffers contribute to increase the accuracy of both URL: https://doi.org/10.1177/10464964211015286.
tasks, translating in overall better system’s accuracy. [5] M. Wozniak, C. Napoli, E. Tramontana, G. Capizzi,
It must be noted that accuracy can be further improved G. Lo Sciuto, R. K. Nowicki, J. T. Starczewski, A mul-
by fine tuning the thresholds required by the instance tiscale image compressor with rbfnn and discrete
segmentation task. A general thumb rule is that, if the wavelet decomposition, in: Proceedings of the Inter-
accuracy is similar between the system using the buffers national Joint Conference on Neural Networks, vol-
and the system not using it, it is possible to improve the ume 2015-September, 2015. doi:10.1109/IJCNN.
performance through such fine tuning. 2015.7280461.
[6] G. Capizzi, S. Coco, G. L. Sciuto, C. Napoli, A new
iterative fir filter design approach using a gaus-
Input C.R. C.R. \w B. I.S. I.S. \w B. Overall %
Kitchen 0.91 1.0 0.89 0.98 0.98
sian approximation, IEEE Signal Processing Letters
Bathroom 0.74 0.81 0.89 1.0 0.81 25 (2018) 1615 – 1619. doi:10.1109/LSP.2018.
Bedroom 0.92 1.0 0.5 0.78 0.78 2866926.
[7] D. Połap, M. Woźniak, C. Napoli, E. Tramontana,
Table 2
R. Damaševičius, Is the colony of ants able to rec-
Performance of the system
ognize graphic objects?, Communications in Com-
puter and Information Science 538 (2015) 376 – 387.
doi:10.1007/978-3-319-24770-0_33.
[8] M. Woźniak, D. Połap, M. Gabryel, R. K. Nowicki,
5. Conclusions C. Napoli, E. Tramontana, Can we process 2d im-
ages using artificial bee colony?, in: Lecture Notes
In this paper we presented a machine learning-powered
in Artificial Intelligence (Subseries of Lecture Notes
solution for privacy enforcement in video data, a data-
in Computer Science), volume 9119, 2015, p. 660 –
driven implementation to safeguard the privacy of any
671. doi:10.1007/978-3-319-19324-3_59.
user that may be forced to spend plenty of hours in videos
[9] A. L. Allen, Dredging up the past: Lifelogging,
and/or in video meetings. Such solution will solve an un-
memory and surveillance, University of Chicago
touched problem that wasn’t formally faced by the field
Law Review 12 (2008) 2825–2830.
in the past years, where our time started to be off-centre
[10] T. Dobrilova, The most astonishing facebook statis-
towards the time spent online.
tics in 2022, 2022. URL: https://techjury.net/blog/
Our implementation presents good performance even in
facebook-statistics/.
presence of noisy or foggy videos, while perform almost
[11] S. Cunningham, M. Masoodian, A. Adams, Pri-
perfectly in the most common scenarios of videos with
vacy issues for online personal photograph collec-
common perspectives extracted from general recordings
tions, Journal of Theoretical and Applied Elec-
using mobile devices. The adaptability of the system to
tronic Commerce Research 5 (2010). doi:10.4067/
change with the needs of different users both for the
S0718-18762010000200003.
objects of interest and the context of interest makes the
[12] R. Xu, N. Baracaldo, J. Joshi, Privacy-preserving ma-
solution we propose a solid step forward in the field of
chine learning: Methods, challenges and directions,
privacy enforcement for video data.
2021. arXiv:2108.04417.
[13] S. Sharma, K. Chen, Image disguising for privacy-
preserving deep learning, in: Proceedings of the
2018 ACM SIGSAC Conference on Computer and
58
Pietro Manganelli Conforti et al. CEUR Workshop Proceedings 53–59
Communications Security, Association for Com- ing a machine learning approach with gan-based
puting Machinery, New York, NY, USA, 2018. URL: data augmentation technique trained using a cus-
https://doi.org/10.1145/3243734.3278511. tom dataset, OBM Neurobiology 6 (2022). doi:10.
[14] G. M. Farinella, C. Napoli, G. Nicotra, S. Riccobene, 21926/obm.neurobiol.2204139.
A context-driven privacy enforcement system for [25] E. Iacobelli, S. Russo, C. Napoli, A machine learning
autonomous media capture devices, Multimedia based real-time application for engagement detec-
Tools and Applications 78 (2019) 14091–14108. URL: tion, in: CEUR Workshop Proceedings, volume
https://doi.org/10.1007/s11042-019-7376-z. 3695, 2023, p. 75 – 84.
[15] P. Speciale, J. L. Schönberger, S. B. Kang, S. N. Sinha, [26] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet
M. Pollefeys, Privacy preserving image-based local- classification with deep convolutional neural net-
ization, CoRR abs/1903.05572 (2019). works (2012).
[16] A. T.-Y. Chen, M. Biglari-Abhari, K. I.-K. Wang, [27] Y. Wu, A. Kirillov, F. Massa, W.-Y. Lo, R. Girshick,
Trusting the computer in computer vision: A Detectron2, https://github.com/facebookresearch/
privacy-affirming framework, in: 2017 IEEE Confer- detectron2, 2019.
ence on Computer Vision and Pattern Recognition [28] T.-Y. Lin, M. Maire, S. Belongie, L. Bourdev, R. Gir-
Workshops (CVPRW), 2017. shick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick,
[17] I. E. Tibermacine, A. Tibermacine, W. Guettala, P. Dollár, Microsoft coco: Common objects in con-
C. Napoli, S. Russo, Enhancing sentiment anal- text, 2015. arXiv:1405.0312.
ysis on seed-iv dataset with vision transformers: [29] RobinReni, House rooms image dataset, 2020.
A comparative study, in: ACM International URL: https://www.kaggle.com/robinreni/
Conference Proceeding Series, 2023, p. 238 – 246. house-rooms-image-dataset.
doi:10.1145/3638985.3639024. [30] A. Buslaev, V. I. Iglovikov, E. Khvedchenya, A. Pari-
[18] V. Ponzi, S. Russo, A. Wajda, R. Brociek, C. Napoli, nov, M. Druzhinin, A. A. Kalinin, Albumentations:
Analysis pre and post covid-19 pandemic rorschach Fast and flexible image augmentations, Informa-
test data of using em algorithms and gmm mod- tion 11 (2020) 125. URL: http://dx.doi.org/10.3390/
els, in: CEUR Workshop Proceedings, volume 3360, info11020125.
2022, p. 55 – 63. [31] R. Girshick, I. Radosavovic, G. Gkioxari, P. Dollár,
[19] A. Alfarano, G. De Magistris, L. Mongelli, S. Russo, K. He, Detectron, 2018. URL: https://github.com/
J. Starczewski, C. Napoli, A novel convmixer trans- facebookresearch/detectron.
former based architecture for violent behavior de- [32] F. Massa, R. Girshick, maskrcnn-benchmark: Fast,
tection, in: Lecture Notes in Computer Science modular reference implementation of Instance
(including subseries Lecture Notes in Artificial In- Segmentation and Object Detection algorithms
telligence and Lecture Notes in Bioinformatics), vol- in PyTorch, https://github.com/facebookresearch/
ume 14126 LNAI, 2023, p. 3 – 16. doi:10.1007/ maskrcnn-benchmark, 2018. Accessed: [Insert date
978-3-031-42508-0_1. here].
[20] E. Iacobelli, V. Ponzi, S. Russo, C. Napoli, Eye- [33] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan,
tracking system with low-end hardware: Devel- S. Belongie, Feature pyramid networks for object
opment and evaluation, Information (Switzerland) detection, 2017. arXiv:1612.03144.
14 (2023). doi:10.3390/info14120644. [34] A. Khvostikov, K. Aderghal, J. Benois-Pineau,
[21] F. Fiani, S. Russo, C. Napoli, An advanced solu- A. Krylov, G. Catheline, 3d cnn-based classifica-
tion based on machine learning for remote emdr tion using smri and md-dti images for alzheimer
therapy, Technologies 11 (2023). doi:10.3390/ disease studies (2018).
technologies11060172. [35] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
[22] V. Ponzi, S. Russo, V. Bianco, C. Napoli, A. Wa- B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer,
jda, Psychoeducative social robots for an healthier R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
lifestyle using artificial intelligence: a case-study, D. Cournapeau, M. Brucher, M. Perrot, E. Duch-
in: CEUR Workshop Proceedings, volume 3118, esnay, Scikit-learn: Machine learning in Python,
2021, p. 26 – 33. Journal of Machine Learning Research 12 (2011)
[23] R. Brociek, G. D. Magistris, F. Cardia, F. Coppa, 2825–2830.
S. Russo, Contagion prevention of covid-19 by
means of touch detection for retail stores, in: CEUR
Workshop Proceedings, volume 3092, 2021, p. 89 –
94.
[24] S. Pepe, S. Tedeschi, N. Brandizzi, S. Russo, L. Ioc-
chi, C. Napoli, Human attention assessment us-
59