Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

CarboFormer: A Lightweight Semantic Segmentation Architecture for Efficient Carbon Dioxide Detection Using Optical Gas Imaging

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CarboFormer claims lightweight transformer segmentation can spot CO2 plumes in thermal video with 84.88% mIoU on controlled releases and 92.98% on livestock emissions, at 84.68 FPS with 5.07M parameters.

desk verdict Worth a serious look for the two new CO2 plume datasets, but the unvalidated mask pipeline puts the accuracy claims on shaky ground. read the letter →

arxiv 2506.05360 v2 pith:GMQDCTJN submitted 2025-05-23 cs.CV

Taminul Islam , Toqi Tahamid Sarker , Mohamed G Embaby , Khaled R Ahmed , Amer AbuGhazaleh This is my paper · ORCID
classification cs.CV
keywords OpticalgasimagingCO2plumesegmentationsemanticlightweighttransformerlow-flowdetectionprecisionlivestockmonitoringCCRdatasetRTA
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

CarboFormer is a compact semantic segmentation model built to find carbon dioxide plumes in thermal video from an optical gas imaging camera, and the paper's central claim is that it does so more accurately than existing lightweight segmentation models while staying fast enough for drones. The authors introduce two datasets: the Controlled Carbon Dioxide Release (CCR) set with flows from 10 to 100 SCCM, and the Real Time Ankom (RTA) set capturing CO2 from dairy-cow rumen fluid experiments. On CCR, CarboFormer reports 84.88% mean IoU and on RTA 92.98%, beating lightweight baselines such as SegFormer-B0 and SegNeXt with only 5.07M parameters and 84.68 FPS. If the claim holds, real-time CO2 monitoring in agriculture and industry could run on small, power-limited devices rather than laboratory instruments.

What carries the argument

The architecture is a four-stage hierarchical encoder with channel widths [32, 64, 160, 256] and spatial reduction ratios [8, 4, 2, 1], using overlap patch merging, Mix-FFN blocks, and attention heads. The decoder combines stage features through a multi-scale harmonic aggregation $F_{out} = \sum_{i=1}^{4} w_i \cdot T_i(F_i)$ with learned weights, and an auxiliary supervision branch at stage S3 adds a loss term $0.4 \mathcal{L}_{aux}$ to the main segmentation loss. This machinery preserves boundary detail in low-flow frames while aggregating global plume context in high-flow frames.

What would settle it

Take the CCR dataset's low-flow sequences (10-30 SCCM) and compare the published ground-truth masks against an independent measurement of plume extent, such as a dense array of calibrated CO2 sensors or manual expert annotation from a different observer; a large disagreement, or a re-evaluation showing CarboFormer's edge over SegFormer-B0 shrinks under the corrected masks, would settle the claim.

Watch

Extended reading notes

Core claim

CarboFormer establishes that a purpose-built lightweight transformer with adaptive hierarchical feature scaling, multi-scale harmonic aggregation, and an auxiliary supervision head can segment CO2 plumes across a 10-100 SCCM flow range more accurately than general-purpose lightweight segmentation networks. The reported numbers are 84.88% mIoU on CCR and 92.98% mIoU on RTA, surpassing SegFormer-B0 (83.36% CCR) and SegNeXt (82.55% CCR) at comparable parameter counts, with 5.07M parameters, 11.39 GFLOPs, and 84.68 FPS. The authors interpret this as evidence that domain-specific design, not just scale, matters for detecting subtle low-flow thermal signatures.

Load-bearing premise

The ground-truth masks come from a semi-automated differential background pipeline with flow-rate-calibrated adaptive thresholding and watershed refinement, and if those masks misrepresent the true plume shapes, especially at the lowest flow rates, every reported mIoU measures agreement with an algorithmic proxy rather than physical ground truth.

Editorial extensions

If this is right

  • At 84.68 FPS with 5.07M parameters, CarboFormer is positioned as deployable on programmable drones for near-real-time CO2 leak surveys.
  • The CCR dataset gives the community a controlled benchmark spanning 10-100 SCCM, where low-flow cases are the hardest and where CarboFormer claims its largest advantage.
  • The RTA results indicate that transfer learning from controlled releases to realistic livestock-emission imagery can work with only 613 training images.
  • The ablation study ties each architectural choice to a specific gain: the auxiliary head adds about 2.4 mIoU, the fourth decoder stage about 4.8 mIoU over a 3-stage decoder, and the third transformer layer about 3.7 mIoU.

Reading between the lines

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

  • If the mask-generation pipeline is biased toward smooth, thresholded shapes, part of CarboFormer's reported advantage may reflect fitting those algorithmic labels rather than the true plume; an independent physical validation would separate the two.
  • The same architecture and loss design should transfer to other mid-wave infrared-absorbing gases such as methane, since the segmentation mechanism is gas-agnostic and only the absorption band and camera change.
  • The paper's 'quantification' claim is really pixel-area segmentation; converting predicted mask area into mass emission rate would require a calibration curve per flow rate, which the paper does not provide but which its datasets make possible.
  • An explicit flow-rate conditioning input, such as regressing or classifying SCCM alongside segmentation, could be a testable extension to sharpen low-flow performance beyond the current auxiliary-head strategy.
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

5 major / 5 minor

Summary. The paper proposes CarboFormer, a lightweight semantic segmentation architecture for CO2 plume segmentation in optical gas imaging, built from a four-stage hierarchical encoder, multi-scale feature integration, and an auxiliary supervision branch. It introduces two datasets: CCR, with controlled CO2 releases at 10-100 SCCM, and RTA, with emissions from dairy cow rumen fluid in vitro experiments. On CCR the model is reported to reach 84.88% mIoU with 5.07M parameters and 84.68 FPS, exceeding SegFormer-B0 (83.36%) and SegNeXt (82.55%); on RTA it reports 92.98% mIoU, also the best among the compared models. Ablations support the three-layer encoder, the four-stage decoder, and the auxiliary head. The ground-truth masks for both datasets come from a semi-automated differential-background pipeline with per-flow-rate adaptive thresholding and watershed refinement, and the paper claims but does not demonstrate validation against physical gas dispersion models. The manuscript acknowledges limited RTA size and potential low-flow improvements, but does not address the unvalidated mask pipeline.

Significance. If the reported results are externally valid, the paper would provide a useful lightweight baseline and two datasets for an underexplored application domain (CO2 optical gas imaging, including livestock emissions). Strengths include five-seed reporting on CCR, a standard MMSegmentation training protocol, a clear efficiency-accuracy comparison, and ablations of the main design choices. However, the central comparison is weakened by the fact that all labels are produced by the same algorithmic pipeline: the reported margins over the nearest baselines are small (1.52 and 2.33 mIoU on CCR, 0.63 mIoU on RTA), so unvalidated label bias could plausibly explain the advantage. The paper does not release code, masks, or independent validation data, which limits reproducibility. The architectural claims themselves are internally consistent, but the benchmark claim requires additional external validation before it can be accepted.

major comments (5)
  1. [Section 3.4] The ground-truth masks for both CCR and RTA are produced entirely by the semi-automated differential-background pipeline described in Section 3.4 (temporal averaging, subtraction, adaptive thresholding calibrated per flow rate, watershed, morphological filtering). The sentence "Final masks undergo validation against physically plausible gas dispersion models [Fig. 1]" cites a figure of example images, not a quantitative validation. No independent measurements of plume geometry, no expert hand-labeled subset, and no comparison to calibrated concentration fields are reported. Because every reported mIoU, including the headline 84.88 on CCR and 92.98 on RTA, measures agreement with these algorithmic labels, the small margins over SegFormer-B0 (1.52 on CCR) and SegNeXt (2.33 on CCR) could reflect bias in the label pipeline rather than architectural superiority. This is load-bearing for the central claim. Please provide independent mask validation and report mask-quality metrics stratified by flow rate, especially 10-30 SCCM where the adaptive threshold is most delicate.
  2. [Table 2 / Section 5.2] The RTA results are reported from a single training run, with no standard deviation or confidence interval, in contrast to the five-seed reporting on CCR. Given the very small dataset (613 images) and the transfer-learning protocol, the 0.63 mIoU advantage over SegFormer-B0 cannot be distinguished from noise. Please report multiple seeds and appropriate significance tests for RTA, or explicitly state that the RTA comparison is preliminary.
  3. [Table 3] The ablation table is internally inconsistent: the row "w/o Auxiliary Head" carries a checkmark in the Aux-H column, and the row "w/ Less Layers (D-2)" carries a checkmark in the D-3 column. As printed, the table cannot be used to determine which configuration was evaluated, undermining the conclusions about decoder depth and auxiliary supervision. Please correct the notation and, ideally, give explicit configuration descriptions rather than checkmarks.
  4. [Section 5.2] The paper claims "statistical significance (p < 0.01)" and says CarboFormer "significantly exceeds" baselines, but the p-value is computed from the five seeds of CarboFormer alone; no pairwise statistical test against SegFormer-B0 or SegNeXt is reported. Without confidence intervals for the baseline mIoU values, the significance claim is unsupported.
  5. [Section 2] The claim that "no published research has utilized semantic segmentation to detect and quantify CO2 emissions from dairy cow rumen fluid" is contradicted by reference [13], which applies OGI and deep learning to rumen fermentation in vitro. Please reconcile this claim and discuss the relationship to the prior work explicitly.
minor comments (5)
  1. [Sections 3.2 and 3.3] The notation "±20 inches" and "±86-90°F" is unusual; the plus-minus sign appears to mean "approximately" rather than a tolerance interval, and "±14.09 PSIA" is likewise unclear. Please rephrase as approximate values or with explicit uncertainties.
  2. [Table 1] The columns "Aux Head" and "Neck" are not defined in the caption or text; please explain what these marks indicate and why they differ across models.
  3. [Figure 3] The decoder diagram labels a block as "Hamburger"; this term is neither defined nor used in the text, and appears to be an editing artifact. Please replace it with the actual module name or remove it.
  4. [Abstract and Section 6] The abstract says the model is designed to "detect and quantify" CO2 emissions, but the paper presents only segmentation; no quantification method or error analysis for emission rates is given. Please either add quantification results or soften the claim.
  5. [Section 6] The limitation paragraph mentions small RTA size and low-flow accuracy, but not the absence of independent validation of the mask pipeline; given that this is the main threat to external validity, it should be acknowledged explicitly.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation-level circularity: the benchmark claims are empirical comparisons against algorithmically generated masks, not self-referential derivations; the remaining concern is external validity of the masks, not circularity.

full rationale

The paper's central claims are empirical benchmarks: CarboFormer achieves 84.88% mIoU on CCR and 92.98% mIoU on RTA with 5.07M parameters. These numbers are obtained by training on image-mask pairs and evaluating with standard mIoU; no equation defines the target quantity in terms of the model's own outputs, and no fitted parameter is renamed as a prediction. The mask-generation pipeline (temporal background averaging, subtraction, adaptive thresholding calibrated per flow rate, watershed, morphological filtering) is a data-construction step that precedes and is shared by all compared methods. A model could learn the pipeline's bias, but that would affect the absolute validity of the labels, not the relative ranking of architectures, and it is not a derivation-level circularity. The paper's self-citations ([13] for the ANKOM protocol and [29] for the GasFormer line of work) are background references; they are not used as a uniqueness theorem, an ansatz smuggled in, or as the justification for the architecture's advantage. The statement 'Final masks undergo validation against physically plausible gas dispersion models [Fig. 1]' is unsupported quantitatively, and the external physical validity of the masks at low flow rates is a genuine limitation; however, that is a correctness/external-validity concern, not circularity. Score 2 reflects the minor self-citation presence and the mild label-generation caveat, not a circular derivation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central empirical claim rests on the validity of the mask-generation pipeline and the camera's CO2-specific sensitivity. No new physical entities are introduced. The main free parameter is the per-flow-rate threshold calibration in mask construction, which is not independently validated.

free parameters (3)
  • Per-flow-rate adaptive threshold in mask generation = not reported (calibrated per flow rate)
    Section 3.4: 'adaptive thresholding calibrated per flow rate' is a free parameter that defines the ground-truth masks; it directly shapes every reported mIoU.
  • Auxiliary loss weight = 0.4
    Section 4.3: the 0.4 weight is hand-set to balance main and auxiliary losses; not derived from a principle.
  • Learning rate / drop path / spatial reduction ratios = 6e-5 / 0.1 / [8,4,2,1]
    Section 5.1 and 4.1: standard hyperparameters chosen by the authors; they affect the final numbers but are not fitted to a benchmark in a way that changes the conclusion.
assumptions (3)
  • domain assumption The FLIR G343 OGI camera response in the 4.2-4.4 µm band corresponds to CO2 plume structure
    Section 3.1 and dataset construction; if the camera visualizes other IR absorption, the masks and model detect the wrong quantity.
  • domain assumption The semi-automated mask pipeline yields accurate plume boundaries
    Section 3.4; all training labels and all evaluation metrics depend on this assumption, which is not independently verified.
  • standard math Train/test splits are independent and representative
    Section 5.1; standard supervised learning assumption, not empirically validated for domain shift across flow rates or camera positions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CarboFormer: A Lightweight Semantic Segmentation Architecture for Efficient Carbon Dioxide Detection Using Optical Gas Imaging." pith.science (2026). https://pith.science/paper/GMQDCTJN

@misc{pith2026250605360,
  author       = {Pith},
  title        = {Pith review of: CarboFormer: A Lightweight Semantic Segmentation Architecture for Efficient Carbon Dioxide Detection Using Optical Gas Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GMQDCTJN}},
  note         = {Machine review of arXiv:2506.05360}
}
abstract

Carbon dioxide (CO$_2$) emissions are critical indicators of both environmental impact and various industrial processes, including livestock management. We introduce CarboFormer, a lightweight semantic segmentation framework for Optical Gas Imaging (OGI), designed to detect and quantify CO$_2$ emissions across diverse applications. Our approach integrates an optimized encoder-decoder architecture with specialized multi-scale feature fusion and auxiliary supervision strategies to effectively model both local details and global relationships in gas plume imagery while achieving competitive accuracy with minimal computational overhead for resource-constrained environments. We contribute two novel datasets: (1) the Controlled Carbon Dioxide Release (CCR) dataset, which simulates gas leaks with systematically varied flow rates (10-100 SCCM), and (2) the Real Time Ankom (RTA) dataset, focusing on emissions from dairy cow rumen fluid in vitro experiments. Extensive evaluations demonstrate that CarboFormer achieves competitive performance with 84.88\% mIoU on CCR and 92.98\% mIoU on RTA, while maintaining computational efficiency with only 5.07M parameters and operating at 84.68 FPS. The model shows particular effectiveness in challenging low-flow scenarios and significantly outperforms other lightweight methods like SegFormer-B0 (83.36\% mIoU on CCR) and SegNeXt (82.55\% mIoU on CCR), making it suitable for real-time monitoring on resource-constrained platforms such as programmable drones. Our work advances both environmental sensing and precision livestock management by providing robust and efficient tools for CO$_2$ emission analysis.

Figures

Figures reproduced from arXiv: 2506.05360 by the authors.

Figure 1
Figure 1. Dataset visualization. (a) Controlled carbon dioxide releases in the CCR dataset at different flow rates (10, 40, 70, and 100 SCCM), showing original images (top) and corresponding segmentation masks (bottom). (b) Real-time CO2 emissions from ANKOM Module in the RTA dataset, demonstrating thermal plume patterns (top) and their ground truth masks (bottom). Additionally, no published research has utilized semantic seg… view at source ↗
Figure 2
Figure 2. Qualitative comparison of segmentation results on the RTA dataset. For each row, we show the input image, ground truth mask, and predictions from different models including DDRNet, Swin-T, DeepLabv3+, SegNeXt, FastSCNN, BiseNetV2, SegFormer￾B0, and our proposed CarboFormer. Our CarboFormer model demonstrates superior boundary preservation and accurate detection of complex CO2 plume morphologies while maintaining com… view at source ↗
Figure 3
Figure 3. Architecture of CarboFormer. Our lightweight architecture combines hierar￾chical feature extraction with specialized multi-scale aggregation optimized for the full spectrum of CO2 emissions (10-100 SCCM). The four-stage encoder (S1-S4) captures both subtle low-rate plume structures and pronounced high-rate thermal signatures through adaptive feature scaling, while the decoder integrates multi-resolution features for… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of segmentation results on the CCR dataset at different carbon dioxide flow rates (10, 70, and 100 SCCM). For each row, we show the input image, ground truth mask, and predictions from different models. Our proposed CarboFormer model demonstrates…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. FUME: Fused Unified Multi-Gas Emission Network for Livestock Rumen Acidosis Detection

    cs.CV 2026-01 conditional novelty 6.0 of 10

    FUME classifies rumen acidosis from CO2/CH4 optical gas images with 98.8% accuracy and 81% mIoU, using 1.28M parameters.

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [13]

    IET Image Processing19(1), e13327 (2025)

    Embaby, M.G., Sarker, T.T., AbuGhazaleh, A., Ahmed, K.R.: Optical gas imaging and deep learning for quantifying enteric methane emissions from rumen fermenta- tion in vitro. IET Image Processing19(1), e13327 (2025)

    work page 2025

  2. [1]

    Scientific Reports15(1), 3655 (2025)

    Alam, G.M.I., Arfin Tanim, S., Sarker, S.K., Watanobe, Y., Islam, R., Mridha, M., Nur, K.: Deep learning model based prediction of vehicle co2 emissions with explainable ai integration for sustainable environment. Scientific Reports15(1), 3655 (2025)

    work page 2025

  3. [2]

    IoT 1, 286–308 (2020)

    Araujo, T., Silva, L., Moreira, A.: Evaluation of low-cost sensors for weather and carbon dioxide monitoring in iot context. IoT 1, 286–308 (2020). https: //doi.org/10.3390/iot1020017

    work page doi:10.3390/iot1020017 2020

  4. [3]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Bai, M., Urtasun, R.: Deep watershed transform for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5221–5229 (2017)

    2017

  5. [4]

    Diagnostics14, 785 (2024)

    Bernasconi, S., Angelucci, A., De Cesari, A., Masotti, A., Pandocchi, M., Vacca, F., Zhao, X., Paganelli, C., Aliverti, A.: Recent technologies for transcutaneous oxygen and carbon dioxide monitoring. Diagnostics14, 785 (2024). https://doi.org/10. 3390/diagnostics14080785

    work page 2024

  6. [5]

    Applied Energy225, 332–345 (2018)

    Chen, B., Harp, D.R., Lin, Y., Keating, E.H., Pawar, R.J.: Geologic co2 sequestra- tion monitoring design: A machine learning and uncertainty quantification based approach. Applied Energy225, 332–345 (2018)

    work page 2018

  7. [6]

    In: ECCV (2018)

    Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: ECCV (2018)

    work page 2018

  8. [7]

    In: Communications in Computer and Information Science, pp

    Christensen, M.S.F.: Leveraging the industrial internet of things (iiot) for real-time co2 monitoring. In: Communications in Computer and Information Science, pp. 35–59 (2025). https://doi.org/10.1007/978-3-031-78572-6_3

    work page doi:10.1007/978-3-031-78572-6_3 2025

Show all 36 references
  1. [8]

    ACM Transactions on Sensor Networks17(2), 1–44 (2021)

    Concas, F., Mineraud, J., Lagerspetz, E., Varjonen, S., Liu, X., Puolamäki, K., Nurmi, P., Tarkoma, S.: Low-cost outdoor air quality monitoring and sensor cali- bration. ACM Transactions on Sensor Networks17(2), 1–44 (2021)

    2021

  2. [9]

    https://github.com/open-mmlab/mmsegmentation (2020)

    Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)

    2020

  3. [10]

    Heliyon (2025)

    Ding, N., Xi, Y., Jiang, W., Li, H., Su, J., Yang, S., Lie, T.T.: State-of-the-art carbon metering: continuous emission monitoring systems for industrial applications. Heliyon (2025). https://doi.org/10.1016/j.heliyon.2025.e42308 CarboFormer: A Lightweight Semantic Segmentatio...

    2025 doi

  4. [11]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy,A.,Beyer,L.,Kolesnikov,A.,Weissenborn,D.,Zhai,X.,Unterthiner,T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  5. [12]

    Sensors24(17), 5675 (2024)

    Dubey, R., Telles, A., Nikkel, J., Cao, C., Gewirtzman, J., Raymond, P.A., Lee, X.: Low-cost co2 ndir sensors: Performance evaluation and calibration using machine learning techniques. Sensors24(17), 5675 (2024)

    2024

  6. [14]

    com/products/g343 (2023), accessed: September 18, 2025

    FLIR Systems, Inc.: FLIR G343 Optical Gas Imaging Camera.https://www.flir. com/products/g343 (2023), accessed: September 18, 2025

    2023

  7. [15]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Fu, A., Hosseini, M.S., Plataniotis, K.N.: Reconsidering co2 emissions from computer vision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2311–2317 (2021)

    2021

  8. [16]

    Reviews in Analytical Chemistry (2023)

    Fu, L., You, S., Li, G., Fan, Z.: Enhancing methane sensing with ndir technology: Current trends and future prospects. Reviews in Analytical Chemistry (2023). https://doi.org/10.1515/revac-2023-0062

    2023 doi

  9. [17]

    arXiv preprint arXiv:2209.08575 (2022)

    Guo, M.H., Lu, C.Z., Hou, Q., Liu, Z., Cheng, M.M., Hu, S.M.: Segnext: Re- thinking convolutional attention design for semantic segmentation. arXiv preprint arXiv:2209.08575 (2022)

    2022 arXiv

  10. [18]

    arXiv preprint arXiv:2503.02910 (2025)

    Guo, W., Du, Y., Du, S.: Langgas: Introducing language in selective zero-shot background subtraction for semi-transparent gas leak detection with a new dataset. arXiv preprint arXiv:2503.02910 (2025)

    2025 arXiv

  11. [19]

    Journal of Animal Science94, 570–570 (2016)

    Hess, P.A., Giraldo, P., Williams, R., Moate, P., Beauchemin, K., Eckard, R.: A novel method for collecting gas produced from the in vitro ankom gas production system. Journal of Animal Science94, 570–570 (2016)

    2016

  12. [20]

    Advances in Neural Information Processing Systems37, 63441–63465 (2024)

    Liang, Y., Cao, Z., Deng, S., Dou, H.X., Deng, L.J.: Fourier-enhanced implicit neural fusion network for multispectral and hyperspectral image fusion. Advances in Neural Information Processing Systems37, 63441–63465 (2024)

    2024

  13. [21]

    arXiv preprint arXiv:2103.14030 (2021)

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030 (2021)

    2021 arXiv

  14. [22]

    Electronics and Communications in Japan (2023).https://doi.org/10.1002/ecj

    Matsubara, H.: An activity recognition system at home based on illuminance sensors. Electronics and Communications in Japan (2023).https://doi.org/10.1002/ecj. 12423

    2023 doi

  15. [23]

    Remote Sensing 11(6), 659 (2019)

    Matvienko, G.G., Sukhanov, A.Y.: Application of neural networks for retrieval of the co2 concentration at aerospace sensing by ipda-dial lidar. Remote Sensing 11(6), 659 (2019)

    2019

  16. [24]

    gov/vital-signs/carbon-dioxide/ (2025), accessed: September 18, 2025

    NASA: Global Climate Change: Vital Signs of the Planet.https://climate.nasa. gov/vital-signs/carbon-dioxide/ (2025), accessed: September 18, 2025

    2025

  17. [25]

    Molecules30(3), 650 (2025)

    Orhan, I.B., Zhao, Y., Babarao, R., Thornton, A.W., Le, T.C.: Machine learning descriptors for co2 capture materials. Molecules30(3), 650 (2025)

    2025

  18. [26]

    IEEE Transactions on Intelligent Transportation Systems (2022)

    Pan, H., Hong, Y., Sun, W., Jia, Y.: Deep dual-resolution networks for real-time and accurate semantic segmentation of traffic scenes. IEEE Transactions on Intelligent Transportation Systems (2022)

    2022

  19. [27]

    arXiv preprint arXiv:1902.04502 (2019)

    Poudel, R.P., Liwicki, S., Cipolla, R.: Fast-scnn: Fast semantic segmentation network. arXiv preprint arXiv:1902.04502 (2019)

    2019 arXiv

  20. [28]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Rangel, J., Schmoll, R., Kroll, A.: On scene flow computation of gas structures with optical gas imaging cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 174–182 (2020) 14 T. Islam et al

    2020

  21. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Sarker, T.T., Embaby, M.G., Ahmed, K.R., AbuGhazaleh, A.: Gasformer: A transformer-based architecture for segmenting methane emissions from livestock in optical gas imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5489–5497 (2024)

    2024

  22. [30]

    Unpublished (2019)

    Srabanti, M.: Design of a non-dispersive infra-red (ndir) based co2 sensor to detect the human respiratory co2. Unpublished (2019)

    2019

  23. [31]

    ACS Sensors 6, 1536–1542 (2021)

    Vafaei, M., Amini, A.: Chamberless ndir co2 sensor robust against environmen- tal fluctuations. ACS Sensors 6, 1536–1542 (2021). https://doi.org/10.1021/ acssensors.0c01863

    2021

  24. [32]

    arXiv preprint arXiv:2105.15203 (2021)

    Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203 (2021)

    2021 arXiv

  25. [33]

    International Journal of Computer Vision pp

    Yu, C., Gao, C., Wang, J., Yu, G., Shen, C., Sang, N.: Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision pp. 1–18 (2021)

    2021

  26. [34]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2881–2890 (2017)

    2017

  27. [35]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhou, T., Wang, W., Konukoglu, E., Van Gool, L.: Rethinking semantic segmenta- tion: A prototype view. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2582–2593 (2022)

    2022

  28. [36]

    Sensors 23(5), 2566 (2023)

    Zhu, H., Xie, W., Li, J., Shi, J., Fu, M., Qian, X., Zhang, H., Wang, K., Chen, G.: Advanced computer vision-based subsea gas leaks monitoring: a comparison of two approaches. Sensors 23(5), 2566 (2023)

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.