Pith. sign in

REVIEW 3 major objections 89 references

Open-source 5G platforms that look protocol-compatible can still measure different timing and I/O harnesses, not the same network property.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 21:02 UTC pith:RLGFBCJH

load-bearing objection Solid methodological hygiene paper for open-source 5G/O-RAN testbeds; the abstract’s claim and numbers cohere, but we only have the abstract (wrong full text was supplied), so causal attribution and matrix grounding stay unverified. the 3 major comments →

arxiv 2603.14661 v3 pith:RLGFBCJH submitted 2026-03-15 cs.NI

AtlasRAN: Timing-Aware Evaluation of Open-source 5G Platforms for Integrated Wireless Testbeds

Ryan Barker , Tolunay Seyfi , Alireza Ebrahimi Dorcheh , Julia Boone , Fatemeh Afghah , Joseph Boccuzzi This is my paper
classification cs.NI
keywords open-source 5GO-RANtiming fidelitydigital twinshardware-in-the-loopclaim-to-capability matrixLDPC offloadreal-time factor
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Open-source 5G and O-RAN test environments expose similar interfaces while differing sharply in timing, buffering, transport, synchronization, and observability. The paper therefore argues that functional compatibility is not timing fidelity, and that studies which appear to measure the same network quantity may instead measure different execution harnesses. AtlasRAN is a timing-aware evaluation framework with two reference architectures (CPU-centric and accelerator/twin) plus a claim-to-capability matrix that tells experimenters what a given platform can credibly claim. A CU–DU uplink load comparison of OpenAirInterface RFSim versus Sionna Research Kit (CUDA LDPC offload) shows goodput collapsing under multi-UE load while CPU/GPU utilization falls and the real-time factor drops below one, indicating host-OS IPC and timing starvation rather than decoder saturation. The practical consequence is that integrated wireless testbeds and digital twins must treat timing discipline, transport path, memory movement, and observability as first-class experimental variables.

Core claim

Environments that expose similar 5G/O-RAN protocol interfaces can preserve very different timing, I/O, synchronization, buffering, transport, and observability behavior; therefore functional compatibility does not equal timing fidelity, and measurement claims must be scoped to the actual execution harness. AtlasRAN supplies the reference architectures and claim-to-capability matrix that make those scopes explicit, grounded by the observed goodput collapse and under-fed accelerator behavior in the OAI RFSim versus Sionna-RK uplink study.

What carries the argument

AtlasRAN itself: two reference architectures (CPU-centric path from software emulation through SDR/HIL to O-RU/OFH; accelerator/twin path from offline modeling through code-realistic twins to real-time AI-RAN) together with a compact claim-to-capability matrix that maps experimental claims onto the timing and observability properties a platform can actually support.

Load-bearing premise

The goodput collapse, falling utilization, and real-time factor below one are caused mainly by host-OS inter-process communication and timing effects, and that one CU–DU uplink scenario is enough to ground the general framework for the wider open-source 5G ecosystem.

What would settle it

Repeat the same multi-UE uplink load experiment with instrumentation that isolates memory bandwidth, scheduler policy, and RFSim model fidelity; if goodput recovers while the accelerated decoder saturates and the real-time factor stays near or above one, the host-OS IPC/timing explanation fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The manuscript (as titled/abstracted) argues that open-source 5G/O-RAN platforms can expose similar protocol interfaces while differing sharply in timing, I/O, synchronization, buffering, transport, and observability; thus functional compatibility is not timing fidelity. It introduces AtlasRAN, a timing-aware evaluation framework with two reference architectures (CPU-centric path: software emulation/SDR-HIL/O-RU-OFH; accelerator/twin path: offline modeling/code-realistic twins/real-time AI-RAN) plus a claim-to-capability matrix. The framework is grounded in a CU–DU uplink load comparison of OpenAirInterface RFSim versus Sionna Research Kit (CUDA LDPC offload retaining the OAI host-OS path). Reported results show goodput collapsing with UE concurrency (OAI 114.59→16.35 Mb/s; Sionna-RK 103.34→16.15 Mb/s), near-ideal fairness, falling CPU/GPU utilization, and RFSim real-time factor <1, interpreted as the accelerated decoder being under-fed by host-OS IPC/timing rather than saturated. The paper concludes that integrated testbeds and digital twins should treat timing discipline, transport path, memory movement, and observability as first-class experimental variables. (Note: the supplied full-text body is an unrelated manuscript, ViDscribe.)

Significance. If the AtlasRAN framing and the causal reading of the OAI/Sionna-RK results hold, the work would improve experimental hygiene in open-source 5G/O-RAN research by making timing fidelity an explicit evaluation dimension rather than an implicit assumption. The concrete goodput, fairness, utilization, and RTF numbers, together with the definitional distinction between interface compatibility and timing fidelity, are useful contributions for the testbed and digital-twin communities. The two reference architectures and claim-to-capability matrix, if fully developed and validated in the correct manuscript body, would give practitioners a practical checklist for deciding what a given platform can credibly measure.

major comments (3)
  1. The supplied full manuscript body is not AtlasRAN; it is an entirely different paper (ViDscribe: Multimodal AI for Customizing Audio Description..., CHI EA ’26 / arXiv:2603.14662). Consequently the methods, experimental setup, figures, tables, claim-to-capability matrix, reference-architecture details, and any statistical or profiling support for the CU–DU uplink study cannot be examined. This mismatch prevents verification of every load-bearing claim beyond the abstract and is fatal to a normal technical review of 2603.14661.
  2. Abstract grounding study: the inference that falling CPU/GPU utilization together with RFSim RTF < 1 under rising UE concurrency shows the accelerated decoder is under-fed by host-OS IPC and timing (rather than scheduling policy, memory bandwidth, RFSim model fidelity, instrumentation overhead, or other bottlenecks) is not isolated by the numbers reported in the abstract alone. Without the missing methods/results sections that would contain profiling of IPC latency, queueing, memory movement, or controlled ablations, the causal attribution remains an interpretation rather than a demonstrated result.
  3. Abstract: a single CU–DU uplink load scenario (OAI RFSim vs Sionna-RK) is presented as grounding for two broad reference architectures and a compact claim-to-capability matrix spanning discrete-event simulators, host-OS emulators, SDR/HIL, O-RU/OFH, digital twins, and accelerator-backed runtimes. Even if the body existed, one scenario would need explicit argument that it is representative enough to underwrite the general matrix; that argument is not visible from the abstract.

Circularity Check

0 steps flagged

No significant circularity: AtlasRAN's claim is empirical/definitional (interface match ≠ timing fidelity), grounded in reported measurements rather than a derivation that reduces to its inputs by construction.

full rationale

AtlasRAN argues that open-source 5G environments can share protocol interfaces while differing in timing, I/O, synchronization, buffering, transport, and observability, so functional compatibility is not timing fidelity. That claim is definitional and observational, not a closed-form derivation. The grounding CU–DU uplink load study reports measured goodput collapse (OAI 114.59→16.35 Mb/s; Sionna-RK 103.34→16.15 Mb/s), near-ideal fairness, falling CPU/GPU utilization, and RFSim real-time factor below unity, then interprets the accelerated decoder as under-fed by host-OS IPC/timing rather than saturated. These are experimental outcomes and a causal reading of them, not fitted parameters renamed as predictions, self-definitional identities, uniqueness theorems imported from the authors, or ansatzes smuggled via self-citation. No equation equates a claimed prediction to its own fit by construction. Any later risk that the claim-to-capability matrix is only illustrated on the same platforms is a scope/validation concern, not circularity of the derivation chain. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

Review is abstract-only for AtlasRAN (supplied full text is a different paper, ViDscribe). Load-bearing premises extractable from the abstract: (1) open-source 5G environments can share protocol interfaces while differing in timing/I/O/sync/buffering/transport/observability; (2) the OAI RFSim vs Sionna-RK CU–DU uplink comparison is a valid grounding case for the general framework; (3) falling utilization plus RTF < 1 implies decoder under-feeding by host-OS IPC rather than other bottlenecks. No free parameters are fitted in the abstract; no new physical entities are postulated. Domain assumptions about what constitutes 'credible measurement' of network properties are implicit.

axioms (3)
  • domain assumption Environments that expose similar 5G/O-RAN protocol interfaces can still differ substantially in timing, I/O, synchronization, buffering, transport, and observability behavior.
    Stated as the opening premise of the abstract; underpins the entire claim that functional compatibility ≠ timing fidelity.
  • ad hoc to paper The CU–DU uplink load comparison of OpenAirInterface RFSim versus Sionna Research Kit (CUDA LDPC offload, retained OAI host-OS path) is representative enough to ground the general AtlasRAN architectures and claim-to-capability matrix.
    The abstract presents this single study as the empirical grounding; generalization to the broader platform taxonomy is assumed rather than derived.
  • ad hoc to paper Falling CPU/GPU utilization together with RFSim real-time factor below unity under rising UE concurrency indicates that the accelerated decoder is under-fed by host-OS IPC and timing effects rather than saturated or limited by another bottleneck.
    Causal reading of the reported trends; alternative explanations (scheduling, memory, instrumentation) are not ruled out in the abstract.
invented entities (1)
  • AtlasRAN (timing-aware evaluation framework with two reference architectures and claim-to-capability matrix) no independent evidence
    purpose: Organize what open-source 5G platforms can credibly measure and force reporting of timing discipline, transport path, memory movement, and observability.
    Named contribution of the paper; a methodological construct rather than a physical entity. Independent evidence would be adoption and successful application by other groups to platforms not used in the original study.

pith-pipeline@v1.1.0-grok45 · 17717 in / 3355 out tokens · 37097 ms · 2026-07-14T21:02:20.900742+00:00 · methodology

0 comments
read the original abstract

Open-source 5G and O-RAN experimentation now spans discrete-event simulators, host-OS emulators, SDR hardware-in-the-loop testbeds, O-RU/Open Fronthaul deployments, wireless digital twins, and accelerator-backed RAN runtimes. These environments may expose similar protocol interfaces while preserving very different timing, I/O, synchronization, buffering, transport, and observability behavior. Thus, studies that appear to measure the same network property may instead measure different execution harnesses: functional compatibility is not timing fidelity. This paper presents AtlasRAN, a timing-aware evaluation framework for deciding what an open-source 5G platform can credibly measure. AtlasRAN provides two reference architectures: a CPU-centric path spanning software emulation, SDR/HIL, and O-RU/OFH execution, and an accelerator/twin path spanning offline modeling, code-realistic twins, and real-time AI-RAN runtimes, plus a compact claim-to-capability matrix. We ground the framework in a CU--DU uplink load study comparing OpenAirInterface RFSim with the Sionna Research Kit, which offloads LDPC decoding to CUDA while retaining much of the surrounding OAI host-OS emulation path. As UE concurrency increases, OAI goodput falls from 114.59 Mb/s at one UE to 16.35 Mb/s in the degraded twelve-UE region, while Sionna-RK falls from 103.34 Mb/s to 16.15 Mb/s. Fairness remains near ideal, CPU/GPU utilization falls with load, and the RFSim real-time factor drops below unity, indicating that the accelerated decoder is under-fed by host-OS inter-process communication and timing effects rather than saturated. AtlasRAN therefore argues that integrated wireless testbeds and digital twins should report timing discipline, transport path, memory movement, and observability as first-class experimental variables.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 2 linked inside Pith

  1. [1]

    Katharina Anderer, Karin Müller, Lukas Strobel, Matthias Wölfel, Jan Niehues, and Kathrin Gerling. 2025. Making Lecture Videos Accessible for Students who are Blind or have Low Vision through AI-Assisted Navigation and Visual Question Answering. InInternational ACM SIGACCESS Conference on Computers and Accessibility (ASSETS)

    2025

  2. [2]

    Ali Selman Aydin, Yu-Jung Ko, Utku Uckun, IV Ramakrishnan, and Vikas Ashok

  3. [3]

    InACM International Conference on Information & Knowledge Management (CIKM)

    Non-Visual Accessibility Assessment of Videos. InACM International Conference on Information & Knowledge Management (CIKM)

  4. [4]

    Aaron Bangor, Philip T Kortum, and James T Miller. 2008. An Empirical Evaluation of the System Usability Scale.Intl. Journal of Human–Computer Interaction(2008)

    2008

  5. [5]

    Made by People, Described by People

    Cynthia L. Bennett, Amy Pavel, Aidan Bryant, Laura J. McCarthy Eberly, Renee Shelby, and Shaun K. Kane. 2025. “Made by People, Described by People”: The Changing Work Practices of Audio Description Professionals.ACM Transactions on Accessible Computing(2025)

    2025

  6. [6]

    Aditya Bodi, Pooyan Fazli, Shasta Ihorn, Yue-Ting Siu, Andrew T Scott, Lothar Narins, Yash Kant, Abhishek Das, and Ilmi Yoon. 2021. Automated Video De- scription for Blind and Low Vision Users. InExtended Abstracts of the Conference on Human Factors in Computing Systems (CHI EA)

    2021

  7. [7]

    Campos, Luiz M

    Virgínia P. Campos, Luiz M. G. Gonçalves, Wesnydy L. Ribeiro, Tiago M. U. Araújo, Thaís G. Do Rego, Pedro H. V. Figueiredo, Suanny F. S. Vieira, Thiago F. S. Costa, Caio C. Moraes, Alexandre C. S. Cruz, Felipe A. Araújo, and Guido L. Souza Filho. 2023. Machine Generation of Audio Description for Blind and Visually Impaired People.ACM Transactions on Acces...

    2023

  8. [8]

    Ruei-Che Chang, Chao-Hsien Ting, Chia-Sheng Hung, Wan-Chen Lee, Liang- Jin Chen, Yu-Tzu Chao, Bing-Yu Chen, and Anhong Guo. 2022. OmniScribe: Authoring Immersive Audio Descriptions for 360°Videos. InACM Symposium on User Interface Software and Technology (UIST)

    2022

  9. [9]

    Maryam Cheema, Hasti Seifi, and Pooyan Fazli. 2025. Describe Now: User-Driven Audio Description for Blind and Low Vision Individuals. InACM Designing Interactive Systems Conference (DIS)

    2025

  10. [10]

    Maryam S Cheema, Sina Elahimanesh, Samuel Martin, Pooyan Fazli, and Hasti Seifi. 2025. DescribePro: Collaborative Audio Description with Human-AI Inter- action. InACM SIGACCESS Conference on Computers and Accessibility (ASSETS)

    2025

  11. [11]

    Agnieszka Chmiel and Iwona Mazur. 2022. A Homogenous or Heterogeneous Audience? Audio description Preferences of Persons with Congenital Blindness, Non-congenital Blindness and Low Vision.Perspectives(2022)

    2022

  12. [12]

    Described and Captioned Media Program. 2012. Audio Description Tip Sheet. (2012). https://dcmp.org/learn/227-audio-description-tip-sheet

    2012

  13. [13]

    Yingqiang Gao, Lukas Fischer, Alexa Lintner, and Sarah Ebling. 2025. Audio Description Generation in the Era of LLMs and VLMs: A Review of Transferable Generative AI Technologies. InFindings of the Association for Computational Linguistics (NAACL)

    2025

  14. [14]

    Ricardo E Gonzalez Penuela, Jazmin Collins, Cynthia Bennett, and Shiri Azenkot

  15. [15]

    InACM SIGCHI Conference on Human Factors in Computing Systems (CHI)

    Investigating Use Cases of AI-Powered Scene Description Applications for Blind and Low Vision People. InACM SIGCHI Conference on Human Factors in Computing Systems (CHI)

  16. [16]

    Mina Huh, Fangyuan Xu, Yi-Hao Peng, Chongyan Chen, Danna Gurari, Eunsol Choi, and Amy Pavel. 2024. Long-Form Answers to Visual Questions from Blind and Low Vision People. InConference on Language Modeling (COLM)

    2024

  17. [17]

    Shasta Ihorn, Yue-Ting Siu, Aditya Bodi, Lothar Narins, Jose M Castanon, Yash Kant, Abhishek Das, Ilmi Yoon, and Pooyan Fazli. 2021. NarrationBot and InfoBot: A Hybrid System for Automated Video Description.arXiv:2111.03994 [cs.HC] (2021). ��� �� ���� ����� ������ ����� ���������� ����� ������ ������� ���� ������������ ������ ������ ��� ����� �����

    Pith/arXiv arXiv 2021

  18. [18]

    Lucy Jiang, Crescentia Jung, Mahika Phutane, Abigale Stangl, and Shiri Azenkot

  19. [19]

    It’s Kind of Context Dependent

    “It’s Kind of Context Dependent”: Understanding Blind and Low Vision People’s Video Accessibility Preferences Across Viewing Scenarios. InACM SIGCHI Conference on Human Factors in Computing Systems (CHI)

  20. [20]

    Lucy Jiang, Mahika Phutane, and Shiri Azenkot. 2023. Beyond Audio Description: Exploring 360°Video Accessibility with Blind and Low Vision Users Through Collaborative Creation. InACM SIGACCESS Conference on Computers and Acces- sibility (ASSETS)

    2023

  21. [21]

    Jiho Kim, Arjun Srinivasan, Nam Wook Kim, and Yea-Seul Kim. 2023. Exploring Chart Question Answering for Blind and Low Vision Users. InConference on Human Factors in Computing Systems (CHI)

    2023

  22. [22]

    Yogesh Kulkarni and Pooyan Fazli. 2025. EgoVITA: Learning to Plan and Verify for Egocentric Video Reasoning.arXiv:2511.18242 [cs.CV](2025)

    Pith/arXiv arXiv 2025

  23. [23]

    Chaoyu Li, Eun Woo Im, and Pooyan Fazli. 2025. VidHalluc: Evaluating Temporal Hallucinations in Multimodal Large Language Models for Video Understanding. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2025

  24. [25]

    InVizWiz Grand Challenge Workshop, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    VideoA11y-40K: A Large-Scale Dataset for Accessible Video Understanding. InVizWiz Grand Challenge Workshop, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  25. [26]

    Chaoyu Li, Sid Padmanabhuni, Maryam S Cheema, Hasti Seifi, and Pooyan Fazli

  26. [27]

    InACM SIGCHI Conference on Human Factors in Computing Systems (CHI)

    VideoA11y: Method and Dataset for Accessible Video Description. InACM SIGCHI Conference on Human Factors in Computing Systems (CHI)

  27. [28]

    3Play Media. 2022. The Ultimate Guide to Audio Description. (2022). https: //www.3playmedia.com/learn/popular-topics/audio-description/

    2022

  28. [29]

    Rosiana Natalie, Ruei-Che Chang, Smitha Sheshadri, Anhong Guo, and Kotaro Hara. 2024. Audio Description Customization. InACM SIGACCESS Conference on Computers and Accessibility (ASSETS)

    2024

  29. [30]

    Rosiana Natalie, Jolene Loh, Huei Suen Tan, Joshua Tseng, Ian Luke Yi-Ren Chan, Ebrima H Jarjue, Hernisa Kacorri, and Kotaro Hara. 2021. The Efficacy of Collaborative Authoring of Video Scene Descriptions. InACM SIGACCESS Conference on Computers and Accessibility (ASSETS)

    2021

  30. [31]

    Rosiana Natalie, Joshua Tseng, Hernisa Kacorri, and Kotaro Hara. 2023. Support- ing Novices Author Audio Descriptions via Automatic Feedback. InACM SIGCHI Conference on Human Factors in Computing Systems (CHI)

    2023

  31. [32]

    Zheng Ning, Brianna L Wimer, Kaiwen Jiang, Keyi Chen, Jerrick Ban, Yapeng Tian, Yuhang Zhao, and Toby Jia-Jun Li. 2024. SPICA: Interactive Video Content Exploration through Augmented Audio Descriptions for Blind or Low-Vision Viewers. InACM SIGACCESS Conference on Human Factors in Computing Systems (CHI)

    2024

  32. [33]

    Amy Pavel, Gabriel Reyes, and Jeffrey P. Bigham. 2020. Rescribe: Authoring and Automatically Editing Audio Descriptions. InACM Symposium on User Interface Software and Technology (UIST)

    2020

  33. [34]

    Charity Pitcher-Cooper, Manali Seth, Benjamin Kao, James M Coughlan, and Ilmi Yoon. 2023. You Described, We Archived: A Rich Audio Description Dataset. Journal on Technology and Persons with Disabilities(2023)

    2023

  34. [35]

    Yotam Sechayk, Ariel Shamir, Amy Pavel, and Takeo Igarashi. 2025. VeasyGuide: Personalized Visual Guidance for Low-vision Learners on Instructor Actions in Presentation Videos. InACM SIGACCESS Conference on Computers and Accessi- bility (ASSETS)

    2025

  35. [36]

    Smith-Kettlewell Eye Research Institute. 2026. YouDescribe: Audio Description for YouTube Videos. https://youdescribe.org/

    2026

  36. [37]

    Yunlong Tang, Junjia Guo, Hang Hua, Susan Liang, Mingqian Feng, Xinyang Li, Rui Mao, Chao Huang, Jing Bi, Zeliang Zhang, Pooyan Fazli, and Chenliang Xu

  37. [38]

    InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Vidcomposition: Can MLLMs analyze compositions in compiled videos?. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  38. [39]

    Tess Van Daele, Akhil Iyer, Yuning Zhang, Jalyn C Derry, Mina Huh, and Amy Pavel. 2024. Making Short-Form Videos Accessible with Hierarchical Video Summaries. InACM SIGCHI Conference on Human Factors in Computing Systems (CHI)

    2024

  39. [40]

    Yujia Wang, Wei Liang, Haikun Huang, Yongqi Zhang, Dingzeyu Li, and Lap-Fai Yu. 2021. Toward Automatic Audio Description Generation for Accessible Videos. InACM SIGCHI Conference on Human Factors in Computing Systems (CHI)

    2021

  40. [41]

    Shuchang Xu, Xiaofu Jin, Wenshuo Zhang, Huamin Qu, and Yukang Yan. 2025. Branch Explorer: Leveraging Branching Narratives to Support Interactive 360 Video Viewing for Blind and Low Vision Users. InACM Symposium on User Interface Software and Technology (UIST)

    2025

  41. [42]

    Diana Yee, Amina Kobenova, Rohan Jhangiani, Piper Stickler, and Sri Kurniawan

  42. [43]

    Better Than Nothing

    “Better Than Nothing” or Not Enough? User-Centered Reflections on AI-Generated Audio Descriptions Across Media Formats. InACM SIGACCESS Conference on Computers and Accessibility (ASSETS)

  43. [44]

    Pooyan Fazli Yogesh Kulkarni. 2026. AVATAR: Reinforcement Learning to See, Hear, and Reason Over Video. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2026

  44. [45]

    Beste F Yuksel, Pooyan Fazli, Umang Mathur, Vaishali Bisht, Soo Jung Kim, Joshua Junhee Lee, Seung Jung Jin, Yue-Ting Siu, Joshua A Miele, and Ilmi Yoon

  45. [46]

    InACM Designing Interactive Systems Conference (DIS)

    Human-in-the-Loop Machine Learning to Increase Video Accessibility for Visually Impaired and Blind Users. InACM Designing Interactive Systems Conference (DIS)

  46. [47]

    Beste F Yuksel, Soo Jung Kim, Seung Jung Jin, Joshua Junhee Lee, Pooyan Fazli, Umang Mathur, Vaishali Bisht, Ilmi Yoon, Yue-Ting Siu, and Joshua A Miele

  47. [48]

    InACM SIGCHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA)

    Increasing Video Accessibility for Visually Impaired Users with Human-in- the-Loop Machine Learning. InACM SIGCHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA). ���������� ���������� �� ��� ����������� ����� ����������� ��� ������� ��������� �� ������ ������ ��� �� ���� ����� ������ ����� ���������� ����� A Appendix: Prompt ...

  48. [49]

    Avoid over-describing - Do not include non-essential visual details

  49. [50]

    Description should not be opinionated unless content de- mands it

  50. [51]

    Choose a level of detail based on plot relevance when de- scribing scenes

  51. [52]

    Description should be informative and conversational, in present tense and third-person omniscient

  52. [53]

    Vocabulary used should ensure accuracy, clarity, and conciseness

    The vocabulary should reflect the predominant lan- guage/accent of the program and should be consistent with the genre and tone of the content while also mindful of the tar- get audience. Vocabulary used should ensure accuracy, clarity, and conciseness

  53. [54]

    Consider historical context and avoid words with negative connotations or bias

  54. [55]

    Pay attention to verbs - Choose vivid verbs over bland ones with adverbs

  55. [56]

    Use pronouns only when clear whom they refer to

  56. [57]

    Use comparisons for shapes and sizes with familiar and glob- ally relevant objects

  57. [58]

    Maintain consistency in word choice, character qualities, and visual elements for all audio descriptions

  58. [59]

    Tone and vocabulary should match the target audience’s age range

  59. [60]

    Ensure no errors in word selection, pronunciation, diction, or enunciation

  60. [61]

    Start with general context, then add details

  61. [62]

    Describe shape, size, texture, or color as appropriate to the content

  62. [63]

    Use first-person narrative for engagement if required to engage the audience

  63. [64]

    Use articles appropriately to introduce or refer to subjects

  64. [65]

    Prefer formal speech over colloquialisms, except where appropriate

  65. [66]

    When introducing new terms, objects, or actions, label them first, and then follow with the definitions

  66. [67]

    Also, do not censor content

    Describe objectively without personal interpretation or com- ment. Also, do not censor content

  67. [68]

    Deliver narration steadily and impersonally (but not monotonously), matching the program’s tone

  68. [69]

    Adjust style for emotion and mood according to the program’s genre

    It can be important to add emotion, excitement, and lightness of touch at different points. Adjust style for emotion and mood according to the program’s genre

  69. [70]

    If it is children’s content, tailor language and pace for chil- dren’s TV, considering audience feedback

  70. [71]

    You should describe what you see

    Do not alter, filter, or exclude content. You should describe what you see. Try to seek simplicity and succinctness in your description

  71. [72]

    Prioritize what is relevant when describing action as to not affect user experience

  72. [73]

    Include location, time, and weather conditions when rele- vant to the scene or plot

  73. [74]

    This is so that the intention of the program is conveyed

    Focus on key content for learning and enjoyment when creating audio descriptions. This is so that the intention of the program is conveyed

  74. [75]

    When describing an instructional video/content, describe the sequence of activities first

  75. [76]

    For a dramatic production, include elements such as style, setting, focus, period, dress, facial features, objects, and aesthet- ics

  76. [77]

    Describe what is most essential for the viewer to know in order to follow, understand, and appreciate the intended learning outcomes of the video/content

  77. [78]

    Audio description should describe characters, locations, time and circumstances, on-screen action, and on-screen informa- tion

  78. [79]

    ��� �� ���� ����� ������ ����� ���������� ����� ������ ������� ���� ������������ ������ ������ ��� ����� �����

    Describe only what a sighted viewer can see. ��� �� ���� ����� ������ ����� ���������� ����� ������ ������� ���� ������������ ������ ������ ��� ����� �����

  79. [80]

    Prioritize factual descriptions of traits like hair, skin, eyes, build, height, age, and visible disabilities

    Describe main and key supporting characters’ visual as- pects relevant to identity and personality. Prioritize factual descriptions of traits like hair, skin, eyes, build, height, age, and visible disabilities. Ensure consistency and avoid singling out characters for specific traits. Use person-first language

  80. [81]

    If unable to confirm or if not established in the plot, do not guess or assume racial, ethnic or gender identity

Showing first 80 references.