The DOI matches its Crossref record; 19 of the 19 references it cites were found in a registry or an index; none of the 15 sources its claims cite is checked for support yet.
Towards Generating Digital LIBRAS Signers on Mobile Devices
Open accessAbstract
Sign language is an effective way to communicate with people who have some degree of hearing impairment. However, mastery of such languages is limited to a relatively small number of people. In this context, the use of assistive technologies is an excellent ally in social inclusion. Approaches based on graphics and deep learning have emerged as an effective non-intrusive way to perform sign language recognition (SLR), translation (SLT), and production (SLP). Pose transfer for sign language reenactment by digital signers is one of the tasks that can be carried out with these techniques. Despite the existence of methods addressing this problem in the literature, few of them tackle the deployment of such models on platforms accessible to real users. Therefore, in this paper, we propose the adaptation and implementation of a deep generative model for Brazilian Sign Language (LIBRAS) production on a mobile device. This effort aims toward the generation of synthetic digital LIBRAS signers which could be used to enhance real-time communication with hearing-impaired people.
Overview
AI snapshot
Written from the paper’s full text.
TL;DR
This application paper adapts the SynLibras conditional VAE-GAN for generating digital LIBRAS signers on Android mobile devices, replacing dense pose heatmaps with MediaPipe stick figures and adding pose normalization. Preliminary on-device results show reconstruction and pose transfer, with each synthesized image taking about 2.8 seconds on a Samsung Exynos 7904 and remaining quality/FPS limitations.
Key findings
- The mobile adaptation changed the pose representation from one 2D dense heatmap per body part to stick figures built from Google MediaPipe keypoints, which “considerably mitigated” the model’s computational cost, and introduced pose normalization between conditioning and target signers.
- The Android application was implemented with Android Studio 2023 and PyTorch 2023, designed to perform pose transfer directly on-device without an internet connection.
- The model was trained on a combination of RWTH-PHOENIX-Weather and SynLibras-Pose totaling approximately 40,000 images resized to 256 × 256 pixels, for approximately 13 epochs with a batch size of 14 images.
- Each image synthesized by the application took an average of 2.8 seconds on a Samsung Exynos 7904 with an Octa-core 1.8 GHz processor; the paper reports that pose transfer is observable but fine details remain unrealistic and a higher FPS rate is needed.
Methods
- Model: a conditional VAE-GAN based on SynLibras, with a Prior module, a discriminator classifying generated images as real or fake, pixel-wise normalization in the decoder, and an equalized learning rate.
- Data: RWTH-PHOENIX-Weather (German sign language, 25 frames per second, nine signers, 210 × 260 pixels) and SynLibras-Pose (LIBRAS, approximately 1133 videos from four signers, each video approximately 200 images at 1024 × 1024); poses were computed with MediaPipe.
- Implementation and evaluation: Android platform using MediaPipe, Android Studio 2023, and PyTorch 2023; trained for approximately 13 epochs with batch size 14; tested reconstruction and pose transfer on mobile, with average synthesis time measured on a Samsung Exynos 7904.
Limitations
- The paper states this is an ongoing, preliminary proof of concept: the model was trained for only approximately 13 epochs, generated images still have unrealistic appearances particularly in fine details, and the FPS rate needs improvement.
- The paper states the model performs “partial pose transfer” and that future work will improve image-generation quality, increase FPS, and employ appearance extraction for unseen appearances.
- Evident from the text: evaluation is limited to the Android platform, no quantitative image-quality metrics or user study are reported, and the datasets involve a small number of signers (nine in RWTH-PHOENIX-Weather and four in SynLibras-Pose) with combined images resized to 256 × 256.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
9 Sources not found · 7 No basis given · 2 Own result · 1 Own interpretation
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“At this stage, there are still some unrealistic appearances, particularly regarding fine details.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 3.3 Reconstruction and Generation of Digital Signers
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“In our mobile application, we have shown the preliminary generation of the digital mobile LIBRAS signer as a proof of concept.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 4 CONCLUSION
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Each image synthesized by the application takes an average of 2.8 seconds to be computed on a Samsung Exynos 7904 with an Octa-core 1,8 GHz processor.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 3.3 Reconstruction and Generation of Digital Signers
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“To mitigate this problem, we have changed the heatmaps by stick figures, constructed from keypoints extracted by a low-weight mobile API, the Google MediaPipe, as illustrated in Figure 2.This adaptation has considerably mitigated the issue of the model's computational cost.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 2.1 Model Adaptation to Mobile Devices
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“As these heatmaps have the input image dimension, the model gets heavier as more body parts are taken into consideration, e.g.eyes, nose, mouth, fingers, etc.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 2.1 Model Adaptation to Mobile Devices
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
None of the sources cited for this claim could be matched to a paper in Crossref, so there is nothing to check it against.
In the paper
“On the other hand, most of the methods that tackle LIBRAS focus on recognition and translation [2,4,7,11,19].”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 1 INTRODUCTION
What the sources say
- Not matchedcites [2]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [4]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [7]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [11]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [19]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
In the paper
“The original pose representation in the SynLibras model was based on one 2D dense heatmap for each body part [23].”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 2.1 Model Adaptation to Mobile Devices
What the sources say
- Not matchedcites [23]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“The transfer of body posture between images is one of the tasks that can be accomplished with these models.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 1 INTRODUCTION
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Particularly, we focus on the generation of digital LIBRAS signers on mobile devices, a problem still not explored in the literature.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 4 CONCLUSION
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
None of the sources cited for this claim could be matched to a paper in Crossref, so there is nothing to check it against.
In the paper
“Deep generative models, such as Generative Adversarial Networks (GANs) [9] and Variational Autoencoders (VAEs) [13], along with more recent methods like Diffusion networks [10], enable various tasks in the domain of images, such as image-to-image translation and image synthesis.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 1 INTRODUCTION
What the sources say
- Not matchedcites [9]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [13]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [10]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
- It may state two separate claims as one. 6
- Probably not a substantial claim. 2
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Nevertheless, it is already possible to observe the pose transfer taking place despite the need for a higher frame-per-second (FPS) rate.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 3.3 Reconstruction and Generation of Digital Signers
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
In the paper
“Silveira et al. [23] introduce the synthesis of LIBRAS in images, yet they did not report their use on mobile devices.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 1 INTRODUCTION
What the sources say
- Not matchedcites [23]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Not matched
In the paper
“Incipient approaches for mobile avatars are found in literature, although they do not focus on sign language [16].”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 1 INTRODUCTION
What the sources say
- Not matchedcites [16]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Not matched
In the paper
“However, the number of people proficient in such sign languages is still small, which is insufficient to address all the issues related to the problem [18].”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 1 INTRODUCTION
What the sources say
- Not matchedcites [18]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Not matched
In the paper
“We presented the mobile adaptation and implementation of SynLibras [23], a model capable of reconstructing input images and performing partial pose transfer.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 4 CONCLUSION
What the sources say
- Not matchedcites [23]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Not matched
In the paper
“Some approaches in the literature perform the task of poseconditioned image synthesis [3,14,15,17,22,[25][26][27].However, none of them are based on LIBRAS databases.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 1 INTRODUCTION
What the sources say
- Not matchedcites [3,14,15,17,22,[25][26][27]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Not matched
In the paper
“Conditional Variational Autoencoders (VAEs) have already been explored in the literature [24] as a way to condition the generation of data (images) in these models.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 2 METHODOLOGY
What the sources say
- Not matchedcites [24]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it.
- Not matched
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“However, such a task might be challenging due to the usual high computational cost of these models, particularly deep generative models for image generation.”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 2.1 Model Adaptation to Mobile Devices
hedgedconditional- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: fails — The paper may give no reason why it holds.
Probably not a substantial claim. The paper may give no reason why it holds.
The paper relies on this without support.
Differences in size and position between the conditioning person and the target person negatively affect the image generation results.
“Differences regarding their size and position affect negatively the image generation results.”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > 1 INTRODUCTION
“recognition (SLR), translation (SLT), and production (SLR)”
§ Towards Generating Digital LIBRAS Signers on Mobile Devices > Abstract
“recognition (SLR), translation (SLT), and production (SLP)”
Citations
Cites
Auto-Encoding Variational Bayes
Kingma, D. P., & Welling, M. (2013) · arXiv · doi:10.48550/arxiv.1312.6114 · Cited by 20,067
Progressive Growing of GANs for Improved Quality, Stability, and Variation
Karras, T., Aila, T., Laine, S., & Lehtinen, J. (2017) · arXiv · doi:10.48550/arxiv.1710.10196 · Cited by 4,727
Sign-to-speech translation using machine-learning-assisted stretchable sensor arrays
Zhou, Z., Chen, K. J., Li, X., Zhang, S., Wu, Y., Zhou, Y., Meng, K., Sun, C., He, Q., Fan, W., et al. (2020) · Nature Electronics · doi:10.1038/s41928-020-0428-6 · Cited by 878
Chan, C., Ginosar, S., Zhou, T., & Efros, A. (2019) · 2019 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00603 · Cited by 737
How Many People Use ASL in the United States? Why Estimates Need Updating
Mitchell, R. E., Young, T. A., Bachleda, B., & Karchmer, M. A. (2006) · Sign Language Studies · doi:10.1353/sls.2006.0019 · Cited by 265
Extensions of the Sign Language Recognition and Translation Corpus RWTH-PHOENIX-Weather
Forster, J., Schmidt, C., Koller, O., Bellgardt, M., & Ney, H. (2014) · Language Resources and Evaluation Conference (LREC) · doi:10.63317/27pdd6xpkah6 · Cited by 137
Almeida, S. G. M., Guimarães, F., & Ramírez, J. A. (2014) · Expert systems with applications · doi:10.1016/j.eswa.2014.05.024 · Cited by 100
Saunders, B., Camgöz, N. C., & Bowden, R. (2022) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52688.2022.00508 · Cited by 83
PoNA: Pose-Guided Non-Local Attention for Human Pose Transfer
Li, K., Zhang, J., Liu, Y., Lai, Y., & Dai, Q. (2020) · IEEE Transactions on Image Processing · doi:10.1109/tip.2020.3029455 · Cited by 59
Generation of Indian sign language by sentence processing and generative adversarial networks
Vasani, N., Autee, P., Kalyani, S., et al. (2020) · International Conference on Intelligent Sustainable Systems (ICISS) · doi:10.1109/iciss49785.2020.9315979 · Cited by 27
Research works printed but not linked: 6
- Mary Jo Davidson. 2006. PAULA: A computer-based sign language tutor for hearing adults. In ITS Workshop.
- Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Nets. In Advances in Neural Information Processing Systems, Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger (Eds.). Vol. 27. Curran Associates, Inc.https://proceedings.neurips.cc/paper_files/paper/2014/file/5ca3e9b122f61f8f06494c97b1afccf3-Paper.pdf
- Jonathan Ho Ajay Jain and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems.
- Jose Jair Alves Mendes Junior, Melissa La Banca Freitas, Sergio Luiz Stevan, and Sergio Francisco Pichorim. 2019. Recognition of Libras Static Alphabet with Myo TM and Multi-Layer Perceptron. In CBEB.
- Ronice Muller de Quadros and Lodenir Becker Karnopp. 2004. Língua brasileira de sinais: estudos linguísticos. Porto Alegre: Artmed (2004).
- Kihyuk Sohn, Honglak Lee, and Xinchen Yan. 2015. Learning Structured Output Representation using Deep Conditional Generative Models. In Advances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett (Eds.). Vol. 28. Curran Associates, Inc.https://proceedings.neurips.cc/paper_files/paper/2015/file/8d55a249e6baa5c06772297520da2051-Paper.pdf
Cited by
Sorted by citations
Costa, V., Tomaszewski, J. P., Pereira, L., Santana, B. S., & Corrêa, G. (2025) · Brazilian Symposium on Multimedia and the Web · doi:10.5753/webmedia.2025.16082 · Cited by 3
“– Silveira et al. [56] further adapted the SynLibras CVAE-GAN for generating digital LIBRAS signers on mobile devices.”
Influential
Metadata
- Authors
- Silveira, W., Alaniz, A., Hurtado, M., Mendonça, L., & de Bem, R.
- Affiliations
- Universidade Federal do Rio Grande (BR)
- Federal University of Rio Grande (FURG), Brazil
- Venue
- Symposium on Virtual and Augmented Reality
- Publisher
- ACM
- Published
- 2023-11-06
- Pages
- 269-273
- OpenAlex
- W4390642483
- Semantic Scholar
- 8f5bf92041872ea8b7f5511826ba42d9db8d2d29
- Cited by
- 1
- References
- 8
- Citations by year
- 2025: 1
- FWCI
- 0.24
- Type
- conference paper
- Language
- en
- Open access
- bronze
- doi.org
- Locations
- Symposium on Virtual and Augmented Reality (SVR) (publishedVersion)
- Funders
- FAPERGS (ARD-21/2551-0000678-1)
- Topics
- Physical Sciences › Computer Science › Human-Computer Interaction › Hand Gesture Recognition Systems
- Social Sciences › Psychology › Developmental and Educational Psychology › Hearing Impairment and Communication
- Physical Sciences › Computer Science › Computer Graphics and Computer-Aided Design › Human Motion and Animation
- Keywords
- Brazilian Sign Language (Libras), sign language production, pose transfer, generative models, mobile deployment, assistive technology, hearing loss, real-time communication
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-17, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-17 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
