Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Safety-oriented sidewalk and road segmentation for smartphone-based assistive navigation

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read For smartphone navigation used by blind pedestrians, picking a segmentation model by average accuracy alone can select a model that more often paints road as walkable sidewalk; the paper shows a slightly less accurate model is safer and fas

desk verdict A useful evaluation recipe for assistive sidewalk segmentation, honestly reported, but the headline architecture comparison is confounded—read the order effects, not the model names. read the letter →

arxiv 2607.21137 v1 pith:LFCWPN4Z submitted 2026-07-23 cs.CV cs.LG

Hakan Calim , Anamaria Dumitrescu , Adarsh Bhandary Panambur , Huzaifa Asif , Andreas Maier This is my paper · ORCID
classification cs.CVcs.LG
keywords assistivenavigationsemanticsegmentationblindandvisuallyimpairedpedestrianssidewalkroad-sidewalkconfusionfalse-safeerrortarget-domainadaptationsmartphonedeployment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that for smartphone navigation apps serving blind and visually impaired pedestrians, a segmentation model must be chosen by more than its average accuracy: what matters is whether the model ever paints road pixels as walkable sidewalk, because that error invites a person to step into traffic. To make that argument, the authors assemble a pedestrian-view dataset, unify the class labels of several external urban datasets, and train five lightweight segmentation architectures with staged fine-tuning, synthetic images, and SAM2 pseudo-labels. They find that accuracy and safety-oriented behavior diverge: the model with the highest offline accuracy also has a higher Road-as-Sidewalk error than a slightly less accurate model, which also runs fastest on a phone. The paper concludes that assistive sidewalk perception should be judged by segmentation accuracy, proxy false-safe behavior, and on-device runtime together, and that real-world benefit still requires validation with end users.

What carries the argument

Road-as-Sidewalk Error Rate — the fraction of road pixels predicted as sidewalk — is the load-bearing safety metric; it turns an invisible false-safe failure mode into a number that can be compared across models. Around it, SENSATION-DS provides the target-domain pedestrian-view dataset with a nine-class navigation taxonomy, and staged target-domain adaptation (fine-tune, then add synthetic or SAM2 pseudo-labeled data) is the training mechanism that lets the authors separate accuracy gains from safety gains.

What would settle it

A BVIP user study, or a wearable-mounted trial, in which the model with the lowest Road-as-Sidewalk Error Rate does not reduce actual road incursions or risky crossings compared with the higher-mIoU model would falsify the claim that this offline metric should drive model selection.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the best-scoring model by mean Intersection over Union is not the safest or the most deployable. UPerNet-MobileNetV3 reaches an mIoU of 0.715 but leaves a Road-as-Sidewalk Error Rate of 0.097, while DeepLabV3Plus-MobileNetV3 has lower mIoU (0.690) but the lowest Road-as-Sidewalk Error Rate (0.079), the highest Road IoU (0.761), and the fastest measured Android runtime (7.383 FPS at 512x384). The paper also reports that synthetic images mainly improve accuracy, whereas SAM2 pseudo-labels more consistently reduce Road-as-Sidewalk errors; applying synthetic images first and SAM2 pseudo-labels second gave the lowest false-safe rate in every

Load-bearing premise

The central claim rests on the premise that SENSATION-DS and the offline Road-as-Sidewalk Error Rate capture real BVIP navigation risk; the paper itself states these are proxy metrics, not validated indicators of safety.

Editorial extensions

If this is right

  • Segmentation benchmarks for assistive navigation should report road-sidewalk confusion metrics, not just mIoU.
  • Model selection for BVIP navigation apps should start from a safety-error budget and then filter for smartphone runtime.
  • SAM2 pseudo-labels are a usable way to lower false-safe road-sidewalk errors without additional human annotation.
  • Applying synthetic images before SAM2 pseudo-labels is a data-only recipe that reduced Road-as-Sidewalk errors across every source-data family tested.
  • Android ONNX benchmarking can identify a viable operating point, here DeepLabV3Plus-MobileNetV3 at 512x384, before moving to user testing.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If Road-as-Sidewalk errors are the binding constraint, an app could combine the fast DeepLabV3Plus model at about 7 FPS with temporal smoothing across video frames to filter flickering boundary errors—an extension the paper does not test.
  • The staged adaptation recipe could be transferred to other false-safe class pairs, such as bike-lane-as-sidewalk or curb-as-sidewalk, since the nine-class taxonomy already collects the relevant classes.
  • A natural next measurement is a BVIP user study comparing wayfinding and road-crossing decisions under DeepLabV3Plus versus UPerNet guidance; if users compensate for segmentation errors, the offline metric gap may not translate into mobility risk.
  • The proxy metric could be validated by mounting a phone at chest height and checking whether predicted walkable masks align with actual step-onto-road events in real pedestrian motion.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a safety-oriented semantic segmentation framework for smartphone-based assistive navigation for blind and visually impaired pedestrians (BVIPs). It introduces SENSATION-DS, a chest-height pedestrian-view dataset of 2,752 image-mask pairs with a nine-class navigation-relevant taxonomy, harmonizes several external urban/sidewalk datasets to this taxonomy, and evaluates five segmentation architectures under a staged target-domain adaptation protocol that uses ControlNet-generated synthetic images and SAM2 pseudo-labels as controlled training extensions. Models are compared offline using mIoU, road/sidewalk IoU, a proposed Road-as-Sidewalk Error Rate (R→S) and Critical False Safe Rate (CFS), and on-phone via Android ONNX benchmarking. The central claims are that mIoU alone is insufficient for model selection, that SAM2 pseudo-labels more consistently reduce R→S false-safe errors, and that DeepLabV3Plus-MobileNetV3 provides the best measured trade-off between accuracy, safety-oriented error, and deployment FPS.

Significance. If the stated findings are supported, the paper makes a useful contribution by drawing attention to road-as-sidewalk false-safe errors as a distinct evaluation criterion for BVIP sidewalk perception, and by providing a pedestrian-view dataset with a navigation-oriented label space. The staged adaptation protocol and the combination of offline safety metrics with Android runtime benchmarking are valuable methodological ingredients. The paper is also commendable for explicitly acknowledging that the safety measures are proxies and that BVIP user studies are needed. However, the headline architecture-level claim — that DeepLabV3Plus-MobileNetV3 achieves the lowest R→S and best deployment trade-off — is not supported by the current experimental design because architecture, source-data family, and stage-3 ordering vary simultaneously across the selected candidates. The contribution would become significant once the comparison is controlled or the claims are appropriately scoped.

major comments (4)
  1. [§4.2, Table 2 and abstract] The abstract and §4.2 state that DeepLabV3Plus-MobileNetV3 'achieved the lowest Road-as-Sidewalk Error Rate (0.079).' This is an architecture-level attribution, but the comparison in Tables 2–3 is confounded: candidate C2 (DeepLabV3Plus-MobileNetV3) differs from C1 (UPerNet-MobileNetV3) in architecture, pretraining source (SideGuide vs Cityscapes), and Stage 3 order (Synthetic→SAM2 vs SAM2→Synthetic) simultaneously. C3, which shares UPerNet-MobileNetV3's architecture with C1 but uses Synthetic→SAM2, reaches R→S = 0.084, nearly matching C2's 0.079. Thus the observed difference could be driven by the ordering or pretraining source rather than by the DeepLabV3Plus architecture. The paper must either present a controlled factorial comparison (e.g., both architectures under identical source and order) or explicitly weaken the claim to a statement about the specific pipeline, not the architect
  2. [§4.1, Table 1] Table 1 presents 'family-wise' results of staged target-domain adaptation but does not identify which architecture(s) produced these numbers. The methods say five architectures were evaluated; if Table 1 refers to a single architecture, that architecture must be named. If the values aggregate over multiple architectures, the aggregation rule is unspecified and the table cannot support family-level conclusions. This omission is load-bearing because the article's general claim about the effects of synthetic augmentation versus SAM2 pseudo-labels on R→S rests entirely on this table.
  3. [§3.5, Tables 1–2] Road-as-Sidewalk Error Rate (R→S) and Critical False Safe Rate (CFS) are never formally defined. The qualitative description in Figure 4 tells the reader that 'pink indicates road pixels predicted as sidewalk,' but the exact numerator and denominator are absent. Without a precise definition (e.g., R→S = |road pixels predicted as sidewalk| / |road pixels|, or another normalization), the central safety metric is not reproducible and the safety interpretation is ambiguous. Both metrics should be defined with equations in §3.5.
  4. [§4.1–4.2, Tables 1–2] The R→S and CFS values are reported as point estimates with no standard deviations, confidence intervals, or statistical tests, while mIoU is reported with standard deviations. The central qualitative claims — that SAM2 pseudo-labels 'more consistently' reduce R→S, and that the best-mIoU model is not the safest — rely on differences as small as 0.0808 vs 0.0857 and 0.079 vs 0.097. Given that training was repeated over five random seeds 'where available,' the authors should report the dispersion of the safety metrics. Without it, the 'consistency' claim cannot be distinguished from seed noise.
minor comments (5)
  1. [Table 2] The table gives candidate labels such as 'Accuracy UPerNet-MobileNetV3' and 'Balanced UPerNet-MobileNetV3' that are not self-explanatory. The reader must cross-reference Table 3 to learn the pretraining source and Stage 3 order. Add columns or a note identifying these settings directly.
  2. [§3.5] Several metrics listed as part of the safety-oriented evaluation — Safe-Walkable Precision, Safe-Walkable Recall, Navigational Risk Error — are not reported in any table. Either include them (perhaps in supplementary material) or remove them from the methods list to avoid expectations of results that are not provided.
  3. [§3.1–3.2] The composition of SENSATION-DS and the harmonization mappings are described only qualitatively. A class-wise mapping table from each source dataset to the nine SENSATION-DS classes would improve transparency and reproducibility.
  4. [Data & Code Availability] The statement says evaluation scripts 'can be made available' but gives no concrete repository or timeline. If possible, provide a public repository URL for the scripts and trained model metadata at submission.
  5. [Abstract] The abstract says 'highest Android runtime at 512×384 (7.383 FPS).' Note that FPS values are device- and resolution-dependent; reporting the specific device would help readers interpret the absolute number.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark with held-out test evaluation; no claimed derivation reduces to its inputs.

full rationale

The paper's claims are empirical measurements, not derived predictions. Models are trained on SENSATION-DS training splits with external source-data families, synthetic images, and SAM2 pseudo-labels, while all reported metrics (mIoU, road/sidewalk IoU, Road-as-Sidewalk Error Rate, FPS) are computed on a held-out test set or device benchmark that was not used for fitting. No parameter is fitted to the metric and then reported as a prediction; the 'safety-oriented' candidate is selected by an explicit safety metric, which is standard model selection rather than circular reasoning. The only self-referential element is the use of the EXACT annotation toolset from the same institution, but that tool is not load-bearing to any scientific claim. The paper explicitly acknowledges that the safety metrics are proxies and that deployment benchmarks do not constitute real-world safety evidence, which further supports that no result is being smuggled in by definition. The skeptical concern about uncontrolled comparisons across architecture, pretraining family, and Stage-3 order in Tables 2-4 is a valid internal-validity limitation for the architecture-level conclusion, but it is a confound, not a circularity: the compared quantities are not equal to each other by construction. Therefore no circular step is present, and the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The central claims rest on dataset representativeness, proxy-metric validity, and pseudo-label quality rather than mathematical derivations. No theoretical free parameters are introduced beyond standard training hyperparameters and hand-chosen evaluation configurations.

free parameters (4)
  • Validation-selected loss weights (Dice/CE) = 0.3/0.7
    §3.3 states the supervised objective combined Dice and cross-entropy with validation-selected weights; these weights affect all reported model results.
  • Candidate selection priorities
    §4.2 selects four candidates by hand (accuracy, safety, balanced, architecture diversity); the compared set is not exhaustive and the ranking depends on this choice.
  • Android benchmark input resolutions = 512x384, 768x576
    §4.4 reports FPS/IoU at two hand-chosen resolutions; the deployment conclusions depend on these operating points.
  • Shared training hyperparameters (AdamW, OneCycle, augmentation)
    §3.3 specifies a common protocol, but exact hyperparameter values are not fully reported; these affect all quantitative outcomes.
assumptions (4)
  • domain assumption SENSATION-DS annotation taxonomy and chest-height viewpoint reflect BVIP mobile navigation needs
    §3.1 frames the whole evaluation around this label space; if the taxonomy or viewpoint is not representative, road/sidewalk metrics lose meaning.
  • domain assumption Road-as-Sidewalk Error Rate and Critical False Safe Rate are valid proxies for false-safe navigation behavior
    §3.5 and §5 define and discuss these metrics, but explicitly state they remain proxy metrics not validated with BVIP users.
  • domain assumption SAM2 pseudo-labels and ControlNet synthetic images are accurate enough to improve training without systematic bias
    §3.4 describes filtering of synthetic samples but reports no quantitative quality thresholds; pseudo-label boundary errors are acknowledged.
  • domain assumption Static offline segmentation masks correlate with real-time smartphone navigation safety
    §4.4 measures on-device IoU and FPS, but not end-to-end navigation safety; the paper itself notes this is not evidence of real-world benefit.
invented entities (3)
  • SENSATION-DS dataset
    purpose: Chest-height pedestrian-view target dataset for training and evaluation
    Introduced in §3.1; raw images accessible only upon request and no external public artifact or independent analysis is provided.
  • Road-as-Sidewalk Error Rate (R→S)
    purpose: Proxy false-safe metric measuring road pixels predicted as sidewalk
    Central safety metric in §3.5 and Tables 1-2, but no formal formula is given and no external validation against BVIP outcomes is provided.
  • Critical False Safe Rate (CFS)
    purpose: Secondary safety metric
    Listed in Tables 1-2 without formal definition; treated as a key safety signal but lacks independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safety-oriented sidewalk and road segmentation for smartphone-based assistive navigation." pith.science (2026). https://pith.science/paper/LFCWPN4Z

@misc{pith2026260721137,
  author       = {Pith},
  title        = {Pith review of: Safety-oriented sidewalk and road segmentation for smartphone-based assistive navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LFCWPN4Z}},
  note         = {Machine review of arXiv:2607.21137}
}
read the original abstract

Independent sidewalk mobility is essential for blind and visually impaired pedestrians (BVIPs), yet smartphone-based assistive navigation requires perception models that distinguish walkable sidewalks from adjacent unsafe regions. This study presents a safety-oriented semantic segmentation framework for future mobile guidance. We introduce SENSATION-DS, a chest-height pedestrian-view dataset with 2,752 image-mask pairs and nine-class navigation-relevant taxonomy. External urban and sidewalk datasets were harmonized to this label space, and five segmentation architectures were evaluated using staged target-domain adaptation with mask-conditioned synthetic images and Segment Anything Model 2 (SAM2) pseudo-labels. Models were assessed using mean Intersection over Union (mIoU), road- and sidewalk-specific metrics, Road-as-Sidewalk Error Rate as a proxy false-safe measure, and Android Open Neural Network Exchange benchmarking. Synthetic augmentation generally improved segmentation accuracy, whereas SAM2 pseudo-labels more consistently reduced Road-as-Sidewalk errors. UPerNet-MobileNetV3 achieved the highest offline mIoU (0.715 +/- 0.006), while DeepLabV3Plus-MobileNetV3 achieved the lowest Road-as-Sidewalk Error Rate (0.079) and highest Android runtime at 512x384 (7.383 FPS). These results show that assistive sidewalk perception should be evaluated jointly by segmentation accuracy, proxy false-safe behavior, and smartphone deployment feasibility, while real-world benefit requires validation with BVIP users. This evaluation supports selecting models that balance accurate perception, conservative error behavior, and practical runtime.

Figures

Figures reproduced from arXiv: 2607.21137 by the authors.

Figure 1
Figure 1. Overview of the proposed safety-oriented sidewalk segmentation framework for [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Training-stage pipeline for staged target-domain adaptation to SENSATION-DS. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Accepted ControlNet-generated synthetic sidewalk-scene examples used for controlled [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative examples of road/sidewalk safety-metric overlays. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 8 canonical work pages

  1. [1]

    Ben Mortenson

    Kishore Seetharaman, Atiya Mahmood, Farinaz Rikhtehgaran, Ghazaleh Akbarnejad, Farrukh Chishtie, Mike Prescott, Alison Chung, and W. Ben Mortenson. Influence of the built environment on community mobility of people living with visual disabilities: a scoping review.Urban, Planning and Transport Research, 12(1), January 2024. URL https://doi.org/10.1080/216...

    arXiv 2024

  2. [2]

    Evaluation of walkability and mobility requirements of visually impaired people in urban spaces.Research in Transportation Business & Management, 40:100592, September

    Tiziana Campisi, Matteo Ignaccolo, Giuseppe Inturri, Giovanni Tesoriere, and Vincenza Torrisi. Evaluation of walkability and mobility requirements of visually impaired people in urban spaces.Research in Transportation Business & Management, 40:100592, September

  3. [3]

    Guti´ errez, Miguel Carrasco, Roberto De Fazio, Andr´ es Trujillo-Le´ on, Paolo Visconti, and Fernando Vidal-Verd´ u

    Ricardo Tachiquin, Ramiro Vel´ azquez, Carolina Del-Valle-Soto, Carlos A. Guti´ errez, Miguel Carrasco, Roberto De Fazio, Andr´ es Trujillo-Le´ on, Paolo Visconti, and Fernando Vidal-Verd´ u. Wearable urban mobility assistive device for visually impaired pedestrians using a smartphone and a tactile-foot interface.Sensors, 21(16):5274, August 2021. URL htt...

    work page doi:10.3390/s21165274 2021

  4. [4]

    DeepNA VI: A deep learning based smartphone navigation assistant for people with visual impairments.Expert Systems 14 with Applications, 212:118720, February 2023

    Bineeth Kuriakose, Raju Shrestha, and Frode Eika Sandnes. DeepNA VI: A deep learning based smartphone navigation assistant for people with visual impairments.Expert Systems 14 with Applications, 212:118720, February 2023. URLhttps://doi.org/10.1016/j.eswa. 2022.118720

    arXiv 2023

  5. [5]

    Bergasa, Eduardo Romera, Weijian Hu, Dongming Sun, Junwei Sun, Ruiqi Cheng, Tianxue Chen, and Elena L´ opez

    Kailun Yang, Kaiwei Wang, Luis M. Bergasa, Eduardo Romera, Weijian Hu, Dongming Sun, Junwei Sun, Ruiqi Cheng, Tianxue Chen, and Elena L´ opez. Unifying terrain aware- ness for the visually impaired through real-time semantic segmentation.Sensors, 18(5): 1506, May 2018. URLhttps://doi.org/10.3390/s18051506

    work page doi:10.3390/s18051506 2018

  6. [6]

    SideGuide: A large-scale sidewalk dataset for guiding impaired people

    Kibaek Park, Youngtaek Oh, Soomin Ham, Kyungdon Joo, Hyokyoung Kim, Hyoyoung Kum, and In So Kweon. SideGuide: A large-scale sidewalk dataset for guiding impaired people. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10022–10029. IEEE, October 2020. URLhttps://doi.org/10.1109/ IROS45743.2020.9340734

    arXiv 2020

  7. [7]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213–3223. IEEE, June 2016. URLhttps: //doi.org/10.1109/CVPR.2016.350

    doi:10.1109/cvpr.2016.350 2016

  8. [8]

    DSC-Net: Enhancing blind road semantic segmentation with visual sensor using a dual-branch swin-cnn architecture.Sensors, 24 (18):6075, September 2024

    Ying Yuan, Yu Du, Yan Ma, and Hejun Lv. DSC-Net: Enhancing blind road semantic segmentation with visual sensor using a dual-branch swin-cnn architecture.Sensors, 24 (18):6075, September 2024. URLhttps://doi.org/10.3390/s24186075

    work page doi:10.3390/s24186075 2024

Show all 45 references
  1. [9]

    Ishan Arefin Hossain, Jareen Anjom, and Rashik Iram Chowdhury

    Md. Ishan Arefin Hossain, Jareen Anjom, and Rashik Iram Chowdhury. Towards walkable footpath detection for the visually impaired on bangladeshi roads with smartphones using deep edge intelligence.Array, 26:100388, July 2025. URLhttps://doi.org/10.1016/ j.array.2025.100388

    2025

  2. [10]

    Self-supervised sidewalk perception using fast video semantic seg- mentation for robotic wheelchairs in smart mobility.Sensors, 22(14):5241, July 2022

    Vishnu Pradeep, Redouane Khemmar, Louis Lecrosnier, Yann Duchemin, Romain Rossi, and Benoit Decoux. Self-supervised sidewalk perception using fast video semantic seg- mentation for robotic wheelchairs in smart mobility.Sensors, 22(14):5241, July 2022. URLhttps://doi.org/10.339...

    2022 doi

  3. [11]

    The ApolloScape dataset for autonomous driving

    Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The ApolloScape dataset for autonomous driving. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Work- shops (CVPR W), pages 1067–10676. IEEE, Jun...

    2018

  4. [12]

    The Mapillary Vistas dataset for semantic understanding of street scenes

    Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The Mapillary Vistas dataset for semantic understanding of street scenes. In2017 IEEE International Conference on Computer Vision (ICCV), pages 5000–5009. IEEE, October

  5. [13]

    Brostow, Julien Fauqueur, and Roberto Cipolla

    Gabriel J. Brostow, Julien Fauqueur, and Roberto Cipolla. Semantic object classes in video: A high-definition ground truth database.Pattern Recognition Letters, 30(2):88–97, January 2009. URLhttps://doi.org/10.1016/j.patrec.2008.04.005

    2009 doi

  6. [14]

    BDD100K: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2633–...

    2020

  7. [15]

    Sagar M. Waghmare, Kimberly Wilber, Dave Hawkey, Xuan Yang, Matthew Wilson, Stephanie Debats, Cattalyya Nuengsigkapian, Astuti Sharma, Lars Pandikow, Huisheng Wang, Hartwig Adam, and Mikhail Sirotenko. SANPO: A scene understanding, acces- sibility and human navigation dataset....

    2025

  8. [16]

    Sidewalk-segmentation dataset

    segmind. Sidewalk-segmentation dataset. Hugging Face dataset, 2023. URLhttps: //huggingface.co/datasets/segments/sidewalk-semantic

    2023

  9. [17]

    Semantic image synthesis via diffusion models, 2022

    Wengang Zhou, Weilun Wang, Jianmin Bao, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Semantic image synthesis via diffusion models, 2022. URLhttps: 15 //arxiv.org/abs/2207.00050

    2022

  10. [18]

    Adding conditional control to text- to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text- to-image diffusion models. In2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3813–3824. IEEE, October 2023. URLhttps://doi.org/10. 1109/ICCV51070.2023.00355

    2023

  11. [19]

    SAM 2: Segment anything in images and videos,

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨ adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junt- ing Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Doll´ ar, and Christoph Fei...

  12. [20]

    Road obstacle video segmentation

    Shyam Nandan Rai, Shyamgopal Karthik, Mariana-Iuliana Georgescu, Barbara Caputo, Carlo Masone, and Zeynep Akata. Road obstacle video segmentation. InPattern Recog- nition, Lecture Notes in Computer Science, pages 186–201. Springer Nature Switzerland,

  13. [21]

    Narasimhan

    Anurag Ghosh, Shen Zheng, Robert Tamburo, Khiem Vuong, Juan Alvarez-Padilla, Hail- iang Zhu, Michael Cardei, Nicholas Dunn, Christoph Mertz, and Srinivasa G. Narasimhan. ROADWork: A dataset and benchmark for learning to recognize, observe, analyze and drive through work zones....

    2025

  14. [22]

    Outdoor naviga- tion for visually impaired based on deep learning

    Saleh Shadi, Saleh Hadi, Mohammad Amin Nazari, and Wolfram Hardt. Outdoor naviga- tion for visually impaired based on deep learning. InProceedings of the 6th International Conference on Actual Problems of System and Software Engineering (APSSE 2019), 2019. URLhttps://ceur-ws.o...

    2019

  15. [23]

    Maxime Bleau, Natalina Martiniello, Nathalie Gingras-Royer, Catherine Tardif-Bernier, and Joseph Paul Nemargut. Exploring the use of smartphone applications during navigation-based tasks for individuals who are blind or who have low vision: future direc- tions and priorities.D...

    2025

  16. [24]

    From SAM to CAMs: Exploring segment anything model for weakly supervised semantic segmentation

    Hyeokjun Kweon and Kuk-Jin Yoon. From SAM to CAMs: Exploring segment anything model for weakly supervised semantic segmentation. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19499–19509. IEEE, June

  17. [25]

    Par- ticipatory design with blind users: A scenario-based approach

    Nuzhah Gooda Sahib, Tony Stockman, Anastasios Tombros, and Oussama Metatla. Par- ticipatory design with blind users: A scenario-based approach. InHuman-Computer Inter- action – INTERACT 2013, Lecture Notes in Computer Science, pages 685–701. Springer Berlin Heidelberg, 2013. U...

    2013 doi

  18. [26]

    Large-scale, longitudinal, hybrid participatory design program to create navigation technology for the blind, 2024

    Daeun Joyce Chung, Muya Guoji, Nina Mindel, Alexis Malkin, Fernando Albertorio, Shane Lowe, Chris McNally, Casandra Xavier, and Paul Ruvolo. Large-scale, longitudinal, hybrid participatory design program to create navigation technology for the blind, 2024. URLhttps://arxiv.org...

    2024 arXiv

  19. [27]

    Bertram, Jennifer Maier, Christian Bergler, Christine Kr¨ oger, J¨ orn Voigt, Katharina Breininger, Robert Klopfleisch, and Andreas Maier

    Christian Marzahl, Marc Aubreville, Christof A. Bertram, Jennifer Maier, Christian Bergler, Christine Kr¨ oger, J¨ orn Voigt, Katharina Breininger, Robert Klopfleisch, and Andreas Maier. EXACT: a collaboration toolset for algorithm-aided annotation of im- ages with annotation ...

    2021 doi

  20. [28]

    Segmentation models pytorch

    Pavel Iakubovskii. Segmentation models pytorch. GitHub repository, 2019. URLhttps: //github.com/qubvel-org/segmentation_models.pytorch

    2019

  21. [29]

    rwightman/pytorch-image-models: v0.8.10dev0 release, 2023

    Ross Wightman, Nathan Raw, Alexander Soare, Aman Arora, Chris Ha, Christoph Reich, Fredo Guan, Jakub Kaczmarzyk, MrT23, Mike, SeeFun, Contrastive, Mohammed Rizin, Hyeongchan Kim, Csaba Kert´ esz, Dushyant Mehta, Guillem Cucurull, Kushajveer Singh, Hankyul, Yuki Tatsunami, Andr...

    2023

  22. [30]

    Encoder-decoder with atrous separable convolution for semantic image segmenta- tion

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmenta- tion. InComputer Vision – ECCV 2018, Lecture Notes in Computer Science, pages 833–851. Springer International Publi...

    2018

  23. [31]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Doll´ ar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 936–944. IEEE, July 2017. URLhttps://doi.org/10.1109/C...

    2017 doi

  24. [32]

    Pyramid attention network for semantic segmentation, 2018

    Hanchao Li, Pengfei Xiong, Jie An, and Lingxue Wang. Pyramid attention network for semantic segmentation, 2018. URLhttps://arxiv.org/abs/1805.10180

    2018 arXiv

  25. [33]

    Unified perceptual parsing for scene understanding

    Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. InComputer Vision – ECCV 2018, Lecture Notes in Computer Science, pages 432–448. Springer International Publishing, 2018. URLhttps: //doi.org/10.1007/978-3-030...

    2018 doi

  26. [34]

    Alvarez, and Ping Luo

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers,

  27. [35]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. ImageNet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, June 2009. URLhttps://doi.org/10. 1109/CVPR.2009.5206848

    2009

  28. [36]

    Andrew Howard, Mark Sandler, Bo Chen, Weijun Wang, Liang-Chieh Chen, Mingx- ing Tan, Grace Chu, Vijay Vasudevan, Yukun Zhu, Ruoming Pang, Hartwig Adam, and Quoc V. Le. Searching for MobileNetV3. In2019 IEEE/CVF International Con- ference on Computer Vision (ICCV), pages 1314–1...

    2019

  29. [37]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  30. [38]

    Smith and Nicholay Topin

    Leslie N. Smith and Nicholay Topin. Super-convergence: very fast training of neural networks using large learning rates. InArtificial Intelligence and Machine Learning for Multi-Domain Operations Applications, page 36. SPIE, May 2019. URLhttps://doi. org/10.1117/12.2520589

    2019 doi

  31. [39]

    Iglovikov, Eugene Khvedchenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A

    Alexander Buslaev, Vladimir I. Iglovikov, Eugene Khvedchenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A. Kalinin. Albumentations: Fast and flexible image aug- mentations.Information, 11(2):125, February 2020. URLhttps://doi.org/10.3390/ info11020125. 17

    2020

  32. [40]

    URLhttps://openreview.net/forum?id=OG18MI5TRL

  33. [2017]

    URLhttps://doi.org/10.1109/ICCV.2017.534

    2017 doi

  34. [2021]

    URLhttps://doi.org/10.1016/j.rtbm.2020.100592

    2020

  35. [2024]

    URLhttps://doi.org/10.1109/CVPR52733.2024.01844

    2024

  36. [2025]

    URLhttps://openreview.net/forum?id=Ha6RTeWMd0

  37. [2026]

    URLhttps://doi.org/10.1007/978-3-032-12840-9_13

    doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.