UCloudNet: A Residual U-Net with Deep Supervision for Cloud Image Segmentation

Hewei Wang; Muhammad Salman Pathan; Shaofan Wang; Soumyabrata Dev; Yee Hui Lee; Yijie Li

REVIEW 3 major objections 4 minor 35 references

UCloudNet: A Residual U-Net with Deep Supervision for Cloud Image Segmentation

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that residual connections in a U-Net encoder plus deep-supervision auxiliary losses deliver both higher segmentation accuracy and faster convergence on sky-camera cloud images than previously published methods.

desk verdict Competent incremental engineering with a broken convergence claim and unmatched baselines — worth a referee but not a headline. read the letter →

arxiv 2501.06440 v1 pith:LGVCHEU7 submitted 2025-01-11 cs.CV eess.IV

Yijie Li , Hewei Wang , Shaofan Wang , Yee Hui Lee , Muhammad Salman Pathan , Soumyabrata Dev This is my paper · ORCID

classification cs.CVeess.IV

keywords cloudimagesegmentationU-NetresidualconnectionsdeepsupervisionauxiliarylossSWINySEGskycamerasbinary

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

The paper proposes UCloudNet, a U-Net variant for segmenting clouds in ground-based sky-camera images. It argues that putting residual connections inside the encoder's convolution blocks and adding two auxiliary deep-supervision output branches yields both higher segmentation accuracy and faster convergence than earlier cloud-segmentation approaches. On the full SWINySEG day-and-night dataset, the best configuration (k=4 with deep supervision and learning-rate decay) reaches precision 0.92, recall 0.94, F-measure 0.93, and error rate 0.06, and the authors report the model converges in fewer than 17,500 iterations. The practical motivation is that sky-camera systems need accurate cloud masks quickly for cloud-cover estimation, so a model that trains faster without losing accuracy is directly useful.

What carries the argument

The central object is the modified encoder 'Double Convolution Block' (DCB): two Conv-BatchNorm-ReLU6 groups with a residual shortcut, applied only on the encoder side of U-Net, alongside two auxiliary deep-supervision loss branches placed at 1/2 and 1/4 output resolutions. The total loss is binary cross-entropy at full resolution plus 0.4 times the 1/2-resolution auxiliary loss and 0.2 times the 1/4-resolution auxiliary loss. The residual connection lets gradients flow through early encoder stages and fuses feature maps, while the auxiliary branches are claimed to speed convergence and regularize training so the final output reaches a low, stable loss within the first 10,000 iterations and converges by about 17,500.

What would settle it

Retrain CloudSegNet and the other Table 1 baselines from scratch on the exact same 8:2 split of SWINySEG used for UCloudNet, with identical preprocessing, optimizer settings, and metric thresholds; if CloudSegNet's F-measure then reaches or exceeds 0.93 with error rate at or below 0.06, the paper's headline accuracy claim would be refuted.

Watch

Extended reading notes

Core claim

The paper reports that a U-Net whose encoder convolution blocks contain residual connections, trained with two auxiliary deep-supervision loss branches at 1/2 and 1/4 resolution (weighted 0.4 and 0.2 in the total loss), outperforms the compared cloud-segmentation methods on SWINySEG. On the combined day-and-night dataset, the k=4 configuration with learning-rate decay and deep supervision reaches precision 0.92, recall 0.94, F-measure 0.93, and error rate 0.06, the best values on all four metrics among the methods in Table 1. The authors also state that the model converges in fewer than 17,500 iterations, and they attribute the gains to better feature aggregation from residual connections and to the regularizing effect of deep supervision during early training.

Load-bearing premise

The load-bearing premise is that the baseline results quoted in Table 1 were obtained under evaluation conditions equivalent to the paper's 8:2 split of SWINySEG; if prior methods were tested on different splits or with different preprocessing, the reported accuracy advantage could come from the comparison setup rather than from UCloudNet's architecture.

Editorial extensions

If this is right

If the reported numbers are correct, UCloudNet (k=4) with deep supervision and learning-rate decay is the best-performing method in the comparison on the full SWINySEG dataset, with F-measure 0.93 and error rate 0.06.
The deep-supervision training strategy is shown to cut the number of iterations needed to converge, making the model more practical for real-time sky-camera systems.
Residual connections in the encoder improve feature aggregation, and because the decoder already has U-Net skip concatenation, the residual shortcut is only needed on the encoder side.
The model converges in fewer than 17,500 iterations (100 epochs at batch size 16), which means it can be retrained quickly when new imagery or updated ground truth becomes available.

Reading between the lines

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

A controlled re-run of the baselines on the same 8:2 split is the natural next test; the paper's Table 1 borrows published numbers, so the size of the reported margin is not yet established under identical conditions.
The auxiliary-loss weighting (0.4 and 0.2) is not ablated beyond the on/off comparison, so one could vary these weights and the branch locations to see how much of the speed gain comes from deep supervision versus residual connections.
The same encoder-residual plus auxiliary-loss recipe could plausibly transfer to multi-class cloud masks or cloud-depth estimation, which the authors list as future work without providing evidence yet.
Because daytime and nighttime models both show gains, the architecture may be robust to illumination changes; testing on sky cameras from other geographic sites would show whether the improvement survives outside the Singapore dataset.

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.

Desk Editor's Note

read the letter

The short version: this is a competent but incremental engineering paper. A residual U-Net with deep supervision applied to cloud segmentation, with code and results on GitHub. The gains over CloudSegNet on SWINySEG are real but modest — F-measure 0.93 vs 0.89 on the full set — and the architecture is a straightforward combination of known components. What the paper does well: proper ablations (k=2 vs k=4, with/without auxiliary loss, lr decay), loss curves that show the auxiliary branches help early training, and full reproducibility (code, dataset, results). The residual connection in the encoder is a sensible choice, and the paper explains why they don't add it to the decoder. These are not new ideas, but they are applied cleanly.

The soft spots are the ones the stress-test flags. The headline claim in the conclusion — 'less than 17500 iterations (100 epochs with batch-size 16)' — does not add up. With 6,768 images and an 8:2 split, the training set is ~5,414 images; at batch size 16, that's ~338 iterations per epoch, so 100 epochs is ~33,800 iterations. The paper gives no early-stopping criterion or reduced dataset that would get you to 17,500. So the 'less training consumption' half of the central claim is unsupported as written. That matters because the abstract sells faster convergence as a key advantage. Either fix the arithmetic or state the actual early-stopping schedule.

Second issue: the Table 1 baseline numbers are all taken from prior publications, not retrained under the same protocol. The paper doesn't say what split or preprocessing those baselines used. If the prior methods were tested on different splits, the accuracy comparison is apples-to-oranges. No error bars or repeated runs either. These are standard concerns for a paper like this; they don't invalidate the accuracy gain, but they do mean the superiority claim is provisional.

The evaluation also selects the best configuration (k=4+aux+lr-decay) on the test set, which inflates the result somewhat. That's a common sin in applied papers, but worth a sentence in the revision.

Who is this for? Someone working on ground-based sky-camera systems who wants a ready-to-use segmentation network with a public implementation. It will not change the field. For a conference workshop or a journal like IEEE GRSL, it's a borderline-acceptable contribution. My recommendation: send it to peer review rather than desk-rejecting. The reproducibility and internal ablations earn it a referee. But the reviewers should push for the iteration-count correction, matched or at least explicitly described baselines, and variance reporting. If those are fixed, it becomes a solid reference point for the cloud-segmentation community.

Referee Report

3 major / 4 minor

Summary. The paper proposes UCloudNet, a residual U-Net with deep supervision for binary cloud image segmentation. The encoder blocks contain residual connections, and two auxiliary loss branches are added at 1/2 and 1/4 resolution to accelerate training. The method is evaluated on the SWINySEG dataset (day, night, and full) and compared with prior cloud-segmentation methods in terms of precision, recall, F-measure, and error rate. The paper also reports ablations over the width multiplier k, learning-rate decay, and the auxiliary-loss branches, and provides a public code repository.

Significance. If the reported accuracy and training-efficiency advantages are substantiated, UCloudNet would be a simple and practical improvement for real-time ground-based cloud imaging. The paper has positive features: it includes ablations for the auxiliary loss and learning-rate decay, reports loss curves, provides a reproducibility link, and evaluates on a publicly available dataset. However, the central claims currently rest on an internally inconsistent iteration count and on baseline numbers whose evaluation protocols are not matched, so the significance cannot be assessed until these issues are resolved.

major comments (3)

[Section 4.1 and Section 5] The statement in the Conclusion that the model 'only needs less than 17500 iterations (100 epochs with batch-size 16) to converge' is arithmetically inconsistent with the training configuration described in Section 4.1. With 6768 total images (6078 daytime + 690 nighttime), an 8:2 split gives approximately 5414 training images; at batch size 16 this is roughly 338 iterations per epoch, so 100 epochs corresponds to about 33,800 iterations, not 17,500. If the authors intend an early-stopping criterion or a smaller effective training set, it is not stated. This undermines the 'less training consumption' claim as written.
[Table 1 and Section 4.1] The quantitative comparison in Table 1 is not a matched evaluation. The baseline numbers appear to be taken from earlier publications, and the paper does not state whether those methods were retrained on the same 8:2 split of SWINySEG, nor does it describe their validation protocols. The dataset labels such as 'SWINySEG (day) (augmented SWIMSEG)' are ambiguous: it is unclear whether the baseline results and the UCloudNet results are on exactly the same test images. No error bars, standard deviations, or repeated-run statistics are reported, so the small F-measure differences (e.g., 0.93 vs. 0.92 on the full dataset) cannot be distinguished from evaluation noise.
[Abstract and Section 3.3] The paper claims that deep supervision 'substantially reduces the training time consumption' and that UCloudNet has 'less training consumption' than previous approaches, but no baseline training time, iteration count, or wall-clock measurement is provided for any competing method. The loss curves in Fig. 5 show the training loss of UCloudNet itself, but they do not compare against a UCloudNet variant without the auxiliary loss, so the specific contribution of deep supervision to convergence speed is not demonstrated.

minor comments (4)

[Equation (1)] The binary cross-entropy formula in Eq. (1) is missing parentheses around the sum; as written, the `-1/N` multiplies only the first term. Please write `L(p,y) = -1/N * sum_i [ y_i log p_i + (1-y_i) log(1-p_i) ]`.
[Section 4.4] The sentence describing Fig. 5 is confusing: 'the loss of the final output converges much faster than the loss of x2-down-sample loss and x4-down-sample loss' seems to contradict the idea that auxiliary losses aid early training. Please clarify what is plotted and what conclusion is intended.
[Figure 3] The qualitative figure caption says results are shown for day-time and night-time columns, but the column counts in the caption (1-6 and 7-12) are not explained in the text; consider labeling each input, ground-truth, and prediction row explicitly.
[Section 4.1] The paper reports training for 100 epochs on a Tesla V100 but does not report the wall-clock training time; adding this would strengthen the 'less training consumption' claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper makes an empirical performance claim on an external benchmark; no prediction reduces by construction to a fitted input or self-citation.

full rationale

The paper's central claim is an empirical comparison on the external SWINySEG dataset, not a derivation. UCloudNet is constructed from standard U-Net, ResNet-style residual connections, and deep supervision, all with fixed architectural choices and hand-set hyperparameters (k, auxiliary loss weights 0.4/0.2, learning rate 0.001, batch size 16). None of these parameters is fitted to the test set, and the reported metrics (precision, recall, F-measure, error rate) are not defined in terms of the architecture's construction. The comparison with prior methods, including CloudSegNet by overlapping authors, is an external benchmark comparison rather than a load-bearing self-citation: the paper does not invoke a uniqueness theorem or prior-work ansatz to justify its design. The auxiliary loss branches are evaluated by loss curves, and the claimed accuracy advantage is contingent on the evaluation protocol, which is a correctness/robustness concern rather than circularity. The 'less than 17500 iterations' statement is arithmetically inconsistent with 100 epochs, batch size 16, and an 8:2 split of 6768 images, but that is an internal consistency error, not a circular-reasoning reduction. No step in the paper reduces by construction to its own inputs, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities or mathematical objects. Its free parameters are standard training hyperparameters and architecture choices that were selected by hand or by a small grid of values. The central claim depends heavily on the comparability of prior baseline numbers, which is an unverified domain assumption.

free parameters (6)

Model width multiplier k = 2 and 4 (best results with k=4)
Controls the number of filters in each block; chosen by hand to vary model capacity.
Auxiliary loss weights = 0.4 and 0.2
Hand-set weights in Eq. (2) for the half-resolution and quarter-resolution auxiliary losses.
Initial learning rate = 0.001
Standard Adam setting used for training.
Learning rate decay gamma = 0.95 per epoch
Exponential decay factor applied after each epoch.
Batch size = 16
Training batch size; affects convergence speed and memory usage.
Number of training epochs = 100
Training duration; the paper claims convergence in less than 17,500 iterations, which corresponds to roughly 100 epochs at batch size 16.

assumptions (5)

domain assumption SWINySEG ground-truth cloud masks are accurate and representative.
The entire evaluation compares predicted masks against these labels; if the labels are noisy, all metrics are affected.
domain assumption Prior baseline scores in Table 1 are comparable because they were obtained under similar experimental conditions.
The paper does not retrain baselines and does not describe their splits, so the comparison assumes compatibility.
domain assumption Residual connections improve feature aggregation in the encoder.
This is the design motivation from ResNet, adopted without independent verification in this specific cloud segmentation setting.
domain assumption Deep supervision with auxiliary losses improves convergence and accuracy.
The paper relies on the known benefit of auxiliary losses, although it also provides an ablation supporting this for its configuration.
standard math Binary cross-entropy is the appropriate objective for binary cloud segmentation.
The loss function in Eq. (1) is the standard choice for pixel-wise binary classification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of UCloudNet: A Residual U-Net with Deep Supervision for Cloud Image Segmentation." pith.science (2026). https://pith.science/paper/LGVCHEU7

@misc{pith2026250106440,
  author       = {Pith},
  title        = {Pith review of: UCloudNet: A Residual U-Net with Deep Supervision for Cloud Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LGVCHEU7}},
  note         = {Machine review of arXiv:2501.06440}
}

read the original abstract

Recent advancements in meteorology involve the use of ground-based sky cameras for cloud observation. Analyzing images from these cameras helps in calculating cloud coverage and understanding atmospheric phenomena. Traditionally, cloud image segmentation relied on conventional computer vision techniques. However, with the advent of deep learning, convolutional neural networks (CNNs) are increasingly applied for this purpose. Despite their effectiveness, CNNs often require many epochs to converge, posing challenges for real-time processing in sky camera systems. In this paper, we introduce a residual U-Net with deep supervision for cloud segmentation which provides better accuracy than previous approaches, and with less training consumption. By utilizing residual connection in encoders of UCloudNet, the feature extraction ability is further improved.

Figures

Figures reproduced from arXiv: 2501.06440 by the authors.

**Figure 1.** The architecture of the UCloudNet model. The procedure between the output of model and the segmentation mask has been omitted in this figure. Conv2d, Bn, ReLU6 Conv2d, Bn, ReLU6 Conv2d, Bn, ReLU6 Conv2d, Bn, ReLU6 Max Pooling Conv2d, Bn, ReLU6 Up Sample Conv2d, Bn, ReLU6 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** The structure of ‘Double Convolution Block’ in encoder, decoder, ‘Down Sample Block’, and ‘Up Sample Block’ (from left to right). 14], and more recent deep learning approaches, as demonstrated in [11,15] multi-label image segmentation. Traditional techniques often rely on color feature analysis, static convolution filters, and pixel gradients. For instance, Dev et al. [13] use principal component analysis (PCA) an… view at source ↗

**Figure 3.** Results of cloud segmentation for day-time (1-6 columns) and night-time (7-12 columns). ‘Down Sample’ blocks, and ‘Up Sample’ blocks. We explain these blocks in the following sections. 3.1. Double Convolution Block (DCB) [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: PR curve of UCloudNet with different training configurations on full SWINySEG ground-based cloud segmentation dataset. perform a threshold with p=0.5 on the sigmoid output, shown in [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Loss curve of the final output and auxiliary outputs. 5. CONCLUSION In this paper, we introduce a residual U-Net with deep supervision for cloud-sky segmentation. We train our model with different configurations on various splits of SWINySEG dataset. Our proposed meth…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 32 canonical work pages

[1]

INTRODUCTION Cloud information analysis plays a crucial role in the field of meteorological research, offering valuable insights into weather patterns and facilitating enhanced forecasting meth- ods. As computer vision and machine learning advance, they have expanded into various interdisciplinary fields such as meteorology estimation [1–3] and weather va...

work page
[2]

RELATED WORKS In addressing the challenge of segmenting sky/cloud images, a variety of methods have been developed. These methods are generally categorized into traditional computer vision tech- niques, as outlined in visual model related studies like [12– arXiv:2501.06440v1 [cs.CV] 11 Jan 2025 C CCC L LL Loss Convolutional Layer Double Convolutional Bloc...

work page Pith review arXiv 2025
[3]

Down Sample

ARCHITECTURE Our UCloudNet is based on the U-Net [26] structure which contains a series of decoders and encoders with channels con- catenation in each stage. To compare with the original U-Net structure, we use a hyper-parameter k to control the param- eters amount and inspired by ResNet, we add residual con- nection in each convolution block in encoder w...

work page
[4]

EXPERIMENTS & RESULTS We conduct experiments under different configurations, in- cluding the size of the model (controlled byk), usage of learn- ing rate decay, and the deep supervision strategy. 4.1. Dataset and Training Configurations The cloud segmentation dataset used in our study was sourced from the Singapore Whole Sky Nychthemeron Image SEG- mentat...

work page 2014
[5]

We train our model with different configurations on various splits of SWINySEG dataset

CONCLUSION In this paper, we introduce a residual U-Net with deep su- pervision for cloud-sky segmentation. We train our model with different configurations on various splits of SWINySEG dataset. Our proposed method achieves better performance as compared to the other methods and our experiments prove that deep supervision with auxiliary loss can gain bet...

work page
[6]

NeSNet: A deep network for estimating near-surface pollutant con- centrations,

P. Dey, B. P. Das, Y . H. Lee, and S. Dev, “NeSNet: A deep network for estimating near-surface pollutant con- centrations,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2023

work page 2023
[7]

Es- timating Solar Irradiance Using Sky Imagers,

S. Dev, F. M. Savoy, Y . H. Lee, and S. Winkler, “Es- timating Solar Irradiance Using Sky Imagers,” Atmo- spheric Measurement Techniques , vol. 12, no. 10, pp. 5417–5429, 2019

work page 2019
[8]

Graph-based Local Climate Classification in Iran,

N. Akrami, K. Ziarati, and S. Dev, “Graph-based Local Climate Classification in Iran,” International Journal of Climatology, vol. 42, no. 3, pp. 1337–1353, 2022

work page 2022

Show all 35 references

[9]

A Data-Driven Approach for Accurate Rainfall Prediction,

S. Manandhar, S. Dev, Y . H. Lee, Y . S. Meng, and S. Winkler, “A Data-Driven Approach for Accurate Rainfall Prediction,” IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no. 11, pp. 9323–9331, 2019

2019
[10]

Day- ahead Forecasts of Air Temperature,

H. Wang, M. S. Pathan, Y . H. Lee, and S. Dev, “Day- ahead Forecasts of Air Temperature,” in 2021 IEEE USNC-URSI Radio Science Meeting (Joint with AP-S Symposium). IEEE, 2021, pp. 94–95

2021
[11]

Holistic and lightweight approach for solar irradiance forecasting,

M. Jain, P. Yadav, and S. Dev, “Holistic and lightweight approach for solar irradiance forecasting,” IEEE Trans- actions on Geoscience and Remote Sensing , 2024

2024
[12]

An extremely-low cost ground-based whole sky imager,

M. Jain, I. Gollini, M. Bertolotto, G. McArdle, and S. Dev, “An extremely-low cost ground-based whole sky imager,” in Proc. IEEE International Geoscience and Remote Sensing Symposium (IGARSS) . IEEE, 2021, pp. 8209–8212

2021
[13]

Design of low-cost, compact and weather-proof whole sky im- agers for high-dynamic-range captures,

S. Dev, F. M. Savoy, Y . H. Lee, and S. Winkler, “Design of low-cost, compact and weather-proof whole sky im- agers for high-dynamic-range captures,” in Proc. IEEE International Geoscience and Remote Sensing Sympo- sium (IGARSS). IEEE, 2015, pp. 5359–5362

2015
[14]

Color-based seg- mentation of sky/cloud images from ground-based cam- eras,

S. Dev, Y . H. Lee, and S. Winkler, “Color-based seg- mentation of sky/cloud images from ground-based cam- eras,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 10, no. 1, pp. 231–242, 2016

2016
[15]

Night- time sky/cloud image segmentation,

S. Dev, F. M. Savoy, Y . H. Lee, and S. Winkler, “Night- time sky/cloud image segmentation,” in Proc. IEEE In- ternational Conference on Image Processing (ICIP) . IEEE, 2017, pp. 345–349

2017
[16]

Cloud- segnet: A deep network for nychthemeron cloud image segmentation,

S. Dev, A. Nautiyal, Y . H. Lee, and S. Winkler, “Cloud- segnet: A deep network for nychthemeron cloud image segmentation,” IEEE Geoscience and Remote Sensing Letters, vol. 16, no. 12, pp. 1814–1818, 2019

2019
[17]

Re- trieving cloud characteristics from ground-based day- time color all-sky images,

C. N. Long, J. M. Sabburg, J. Calbó, and D. Pagès, “Re- trieving cloud characteristics from ground-based day- time color all-sky images,” Journal of Atmospheric and Oceanic Technology, vol. 23, no. 5, pp. 633–652, 2006

2006
[18]

Systematic study of color spaces and components for the segmentation of sky/cloud images,

S. Dev, Y . H. Lee, and S. Winkler, “Systematic study of color spaces and components for the segmentation of sky/cloud images,” in Proc. IEEE International Confer- ence on Image Processing (ICIP) . IEEE, 2014, pp. 5102–5106

2014
[19]

An auto- matic groundbased cloud detection method based on lo- cal threshold interpolation,

J. Yang, W. Lv, Y . Ma, W. Yao, and Q. Li, “An auto- matic groundbased cloud detection method based on lo- cal threshold interpolation,” Acta Meteorologica Sinica, vol. 68, no. 6, pp. 1007–1017, 2010

2010
[20]

Multi-label cloud segmentation using a deep network,

S. Dev, S. Manandhar, Y . H. Lee, and S. Winkler, “Multi-label cloud segmentation using a deep network,” in 2019 USNC-URSI Radio Science Meeting (Joint with AP-S Symposium). IEEE, 2019, pp. 113–114

2019
[21]

AMDCNet: An attentional multi-directional convolutional network for stereo matching,

H. Wang, Y . Li, S. Xi, S. Wang, M. S. Pathan, and S. Dev, “AMDCNet: An attentional multi-directional convolutional network for stereo matching,” Displays, vol. 74, p. 102243, 2022

2022
[22]

DMCNet: Diversified model combination network for understanding engage- ment from video screengrabs,

S. Batra, H. Wang, A. Nag, P. Brodeur, M. Check- ley, A. Klinkert, and S. Dev, “DMCNet: Diversified model combination network for understanding engage- ment from video screengrabs,” Systems and Soft Com- puting, vol. 4, p. 200039, 2022

2022
[23]

Stereo Matching Based on Visual Sensitive Information,

H. Wang, M. S. Pathan, and S. Dev, “Stereo Matching Based on Visual Sensitive Information,” in 2021 6th In- ternational Conference on Image, Vision and Comput- ing (ICIVC), 2021, pp. 312–316

2021
[24]

SYGNet: A SVD-YOLO based GhostNet for Real-time Driving Scene Parsing,

H. Wang, B. Zhu, Y . Li, K. Gong, Z. Wen, S. Wang, and S. Dev, “SYGNet: A SVD-YOLO based GhostNet for Real-time Driving Scene Parsing,” in 2022 IEEE Inter- national Conference on Image Processing (ICIP), 2022, pp. 2701–2705

2022
[25]

AbHE: All Attention-Based Homography Estimation,

M. Huo, Z. Zhang, X. Ren, X. Yang, and C. Ye, “AbHE: All Attention-Based Homography Estimation,” IEEE Transactions on Instrumentation and Measure- ment, vol. 73, pp. 1–11, 2024

2024
[26]

Joint Pedestrian Trajectory Prediction through Posterior Sampling,

H. Lin, Y . Wang, M. Huo, C. Peng, Z. Liu, and M. Tomizuka, “Joint Pedestrian Trajectory Prediction through Posterior Sampling,” 2024 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS), pp. 5672–5679, 2024

2024
[27]

Human-oriented Repre- sentation Learning for Robotic Manipulation,

M. Huo, M. Ding, C. Xu, T. Tian, X. Zhu, Y . Mu, L. Sun, M. Tomizuka, and W. Zhan, “Human-oriented Repre- sentation Learning for Robotic Manipulation,” ArXiv, vol. abs/2310.03023, 2023

2023 arXiv
[28]

MENTOR: Multi-level Self-supervised Learn- ing for Multimodal Recommendation,

J. Xu, Z. Chen, S. Yang, J. Li, H. Wang, and E. C.-H. Ngai, “MENTOR: Multi-level Self-supervised Learn- ing for Multimodal Recommendation,” arXiv preprint arXiv:2402.19407, 2024

2024 arXiv
[29]

VGRISys: A Vision-Guided Robotic Intelligent System for Au- tonomous Instrument Calibration*,

Z. Li, H. Wang, Y . Li, S. Dev, and G. Zuo, “VGRISys: A Vision-Guided Robotic Intelligent System for Au- tonomous Instrument Calibration*,” in2023 IEEE Inter- national Conference on Robotics and Biomimetics (RO- BIO), 2023, pp. 1–6

2023
[30]

DAANet: Dual Attention Aggregating Network for Salient Object Detection,

Y . Li, H. Wang, Z. Li, S. Wang, S. Dev, and G. Zuo, “DAANet: Dual Attention Aggregating Network for Salient Object Detection,” in IEEE International Con- ference on Robotics and Biomimetics (ROBIO) . IEEE, 2023, pp. 1–7

2023
[31]

U-net: Con- volutional networks for biomedical image segmenta- tion,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Con- volutional networks for biomedical image segmenta- tion,” in Proc. International Conference on Medical image computing and computer-assisted intervention . Springer, 2015, pp. 234–241

2015
[32]

A hybrid thresholding al- gorithm for cloud detection on ground-based color im- ages,

Q. Li, W. Lu, and J. Yang, “A hybrid thresholding al- gorithm for cloud detection on ground-based color im- ages,” Journal of atmospheric and oceanic technology , vol. 28, no. 10, pp. 1286–1296, 2011

2011
[33]

A simple method for the assessment of the cloud cover state in high-latitude regions by a ground-based digital camera,

M. P. Souza-Echer, E. B. Pereira, L. Bins, and M. An- drade, “A simple method for the assessment of the cloud cover state in high-latitude regions by a ground-based digital camera,” Journal of Atmospheric and Oceanic Technology, vol. 23, no. 3, pp. 437–447, 2006

2006
[34]

Image edge detecting based on gap statistic model and relative en- tropy,

Q. Yang, L. Tang, W. Dong, and Y . Sun, “Image edge detecting based on gap statistic model and relative en- tropy,” inProc. Sixth International Conference on Fuzzy Systems and Knowledge Discovery, vol. 5. IEEE, 2009, pp. 384–387

2009
[35]

Ground- based detection of nighttime clouds above Manila ob- servatory (14.64° n, 121.07° e) using a digital camera,

G. F. B. Gacal, C. Antioquia, and N. Lagrosas, “Ground- based detection of nighttime clouds above Manila ob- servatory (14.64° n, 121.07° e) using a digital camera,” Applied Optics, vol. 55, no. 22, pp. 6040–6045, 2016

2016

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

INTRODUCTION Cloud information analysis plays a crucial role in the field of meteorological research, offering valuable insights into weather patterns and facilitating enhanced forecasting meth- ods. As computer vision and machine learning advance, they have expanded into various interdisciplinary fields such as meteorology estimation [1–3] and weather va...

work page

[2] [2]

RELATED WORKS In addressing the challenge of segmenting sky/cloud images, a variety of methods have been developed. These methods are generally categorized into traditional computer vision tech- niques, as outlined in visual model related studies like [12– arXiv:2501.06440v1 [cs.CV] 11 Jan 2025 C CCC L LL Loss Convolutional Layer Double Convolutional Bloc...

work page Pith review arXiv 2025

[3] [3]

Down Sample

ARCHITECTURE Our UCloudNet is based on the U-Net [26] structure which contains a series of decoders and encoders with channels con- catenation in each stage. To compare with the original U-Net structure, we use a hyper-parameter k to control the param- eters amount and inspired by ResNet, we add residual con- nection in each convolution block in encoder w...

work page

[4] [4]

EXPERIMENTS & RESULTS We conduct experiments under different configurations, in- cluding the size of the model (controlled byk), usage of learn- ing rate decay, and the deep supervision strategy. 4.1. Dataset and Training Configurations The cloud segmentation dataset used in our study was sourced from the Singapore Whole Sky Nychthemeron Image SEG- mentat...

work page 2014

[5] [5]

We train our model with different configurations on various splits of SWINySEG dataset

CONCLUSION In this paper, we introduce a residual U-Net with deep su- pervision for cloud-sky segmentation. We train our model with different configurations on various splits of SWINySEG dataset. Our proposed method achieves better performance as compared to the other methods and our experiments prove that deep supervision with auxiliary loss can gain bet...

work page

[6] [6]

NeSNet: A deep network for estimating near-surface pollutant con- centrations,

P. Dey, B. P. Das, Y . H. Lee, and S. Dev, “NeSNet: A deep network for estimating near-surface pollutant con- centrations,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2023

work page 2023

[7] [7]

Es- timating Solar Irradiance Using Sky Imagers,

S. Dev, F. M. Savoy, Y . H. Lee, and S. Winkler, “Es- timating Solar Irradiance Using Sky Imagers,” Atmo- spheric Measurement Techniques , vol. 12, no. 10, pp. 5417–5429, 2019

work page 2019

[8] [8]

Graph-based Local Climate Classification in Iran,

N. Akrami, K. Ziarati, and S. Dev, “Graph-based Local Climate Classification in Iran,” International Journal of Climatology, vol. 42, no. 3, pp. 1337–1353, 2022

work page 2022