EA-RMENet -- Path Loss Prediction in Urban Environments using Deep Learning

Conor Brennan (DCU School of Electronic Engineering); Jonathan O'Shea (DCU School of Electronic Engineering)

REVIEW 3 major objections 6 minor 16 references

EA-RMENet -- Path Loss Prediction in Urban Environments using Deep Learning

T0 review · 3 major / 6 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 A U-Net with an EfficientNet encoder, attention-gated skip connections, and multi-scale pooling predicts urban radio maps from building geometry alone, reaching 0.033 RMSE and 0.022 seconds per sample.

desk verdict A solid, incremental U-Net variant for radio map estimation with an external challenge benchmark, but the real-world claim rests on simulated data and single-run metrics with no code released. read the letter →

arxiv 2607.16449 v1 pith:P3UDAN2V submitted 2026-07-17 cs.LG

Jonathan O'Shea (DCU School of Electronic Engineering) , Conor Brennan (DCU School of Electronic Engineering) This is my paper · ORCID

classification cs.LG

keywords pathlosspredictionradiomapestimationU-NetEfficientNetattentiongatesatrousspatialpyramidpoolingdataaugmentationurbanpropagation

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 a single image-to-image deep network can replace slow deterministic ray tracing for urban path-loss prediction, producing a full radio map in about 0.022 seconds with accuracy competitive with published models. The proposed EA-RMENet is a U-Net-style encoder-decoder whose encoder is an EfficientNetB5 convolutional network, whose skip connections include learned attention gates, and whose bottleneck uses atrous spatial pyramid pooling to capture multi-scale propagation patterns. On the simulated RadioMapSeer3D dataset the model reports a test RMSE of 0.0334 in normalized path-loss units, and on an independent 2023 radio-map prediction challenge it reports 0.0406, ranking third among submissions. The paper's clearest practical finding is that simple geometric data augmentation is the highest-leverage training enhancement, lowering RMSE by 14.6 percent relative to the same architecture without augmentation, while adding positional grid anchors lowers it by only 3.83 percent. A sympathetic reader would take the paper as evidence that attention-gating and multi-scale pooling are useful additions for image-based coverage prediction, while augmentation is what makes the model generalize.

What carries the argument

The central machinery is the EA-RMENet architecture, a U-Net with three coordinated parts. The EfficientNetB5 encoder is a compound-scaled convolutional network—scaling depth, width, and resolution together—that extracts five hierarchical feature maps while keeping the computational budget low. Attention-gated skip connections weight each encoder feature map by a learned per-location coefficient in [0,1] before fusion with decoder features, suppressing irrelevant information instead of concatenating everything. The ASPP bottleneck applies parallel dilated convolutions with dilation rates 1-4 plus a global-pooling branch, so the decoder receives both fine local detail and broad context such a

What would settle it

Collect drive-test received-power measurements on a real urban grid with known building footprints, heights, and transmitter location, feed the same building/height/transmitter image input to the trained EA-RMENet/DA model, and compute RMSE in dB against the measured map. If the measured RMSE is much larger than the roughly 1.2 dB implied by the 0.0334 normalized RMSE, the central accuracy claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that combining an EfficientNetB5 encoder with attention-gated skip connections and an ASPP bottleneck yields a radio-map estimator that is both accurate and fast enough for practical planning loops. On RadioMapSeer3D, EA-RMENet/DA achieves a test RMSE of 0.0334 on maps normalized to [0,1], corresponding to roughly 1.2 dB over the dataset's dynamic range, with an inference time of 0.0217 seconds per sample. The same model, when submitted to an independent 2023 radio-map prediction challenge with denser, more complex maps, scores 0.0406 and ranks third, a degradation shared by all submissions and attributed to distribution shift. Included ablations show that re

Load-bearing premise

The load-bearing premise is that the ray-traced simulated maps used for training faithfully represent real urban propagation; the paper uses no measured field data, so if the simulator's building geometry, materials, or antenna assumptions do not match reality, the reported RMSEs will not transfer to actual network planning.

Editorial extensions

If this is right

If the reported accuracy holds on real maps, wireless network planners can evaluate dozens of transmitter placements per second instead of waiting minutes or hours for ray tracing, making iterative coverage planning practical.
Because data augmentation on geometric symmetries improved RMSE by 14.6 percent, training-time transformation of radio-map inputs is a cheap, high-impact tool for future radio map estimation models.
Attention-gated skip connections are presented as learned filters that suppress irrelevant encoder features during upsampling, while the ASPP bottleneck provides the multi-scale context needed for shadowing and reflection patterns.
The 0.0406 challenge score on denser, unfamiliar maps suggests the network transfers across urban layouts, though with degradation; the paper states its next step as integrating higher-resolution feature encoding strategies.
EA-RMENet/DA produces a 256x256 radio map in 0.022 seconds, fast enough to sit inside an optimization loop that searches over antenna locations.

Reading between the lines

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

A testable extension the paper does not run: ablate the attention gates and ASPP separately. The paper only ablates augmentation and grid anchors, so the marginal contribution of each architectural component remains unmeasured; a two-by-two ablation would say which part actually carries the accuracy.
The paper's normalized RMSE of 0.0334 corresponds to about 1.2 dB if error is uniform across the map, but path-loss planning margins usually depend on where the errors occur, especially near building edges. The paper does not report spatial error maps, so per-location error behavior is an open question that matters more than the scalar RMSE.
Because the challenge's denser maps caused a 0.0072 RMSE degradation, the model's practical ceiling may be set by how much high-resolution geometric detail survives encoder downsampling; the two higher-ranked systems emphasized high-resolution pathways, suggesting the next experiment is a higher-resolution bottleneck with the same augmentation.
If the simulator's clipping rule at -111 dB masks weak-signal regions, the model is never trained on very deep shadow zones; real deployments with extreme shadowing could expose errors that the metric hides.

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

This is a competent engineering paper: a U-Net with an EfficientNetB5 encoder, attention-gated skip connections, and ASPP, evaluated on RadioMapSeer3D and the ICASSP 2023 challenge. The specific combination of components is not in the prior work, and the external challenge leaderboard gives the accuracy numbers a fair, non-circular benchmark. The data augmentation ablation is the cleanest result — a 14.6% relative RMSE gain over baseline, which is believable and useful. Inference time of 0.022 s/sample is also practical. The authors are honest about the distribution shift on the challenge set.

The soft spots are real but addressable. The central claim about "real-world RME" is not supported: every evaluation uses ray-traced simulated maps, both internal and challenge. Simulator artifacts — isotropic antennas, simplified geometry, hard clipping — may not transfer to field measurements. The authors acknowledge the simulation-to-simulation shift but not this deeper limitation. Second, all accuracy numbers are single-run point estimates on one fixed split. No error bars, no repeated seeds, so the 0.0334 versus 0.0382 gap between DA and GA variants could be partly noise. Third, code and checkpoints are not released, and key implementation details live in thesis appendices that are not in the arXiv post. That makes the headline numbers unverifiable as-is.

None of this is fatal to the core result. The architecture makes sense, the external benchmark is a plus, and the writing is clear. The overclaim is in the framing, not the numbers. If the authors release code, add repeated-seed intervals, and either temper the real-world language or validate on measured data, this becomes a citable reference for U-Net-style RME. As it stands, it is a decent workshop or conference paper, not a definitive study.

I would send it to peer review rather than desk-reject — the engineering is coherent and the leaderboard comparison is valuable to a small but active community — but I would ask for the reproducibility fixes before acceptance.

Referee Report

3 major / 6 minor

Summary. The paper proposes EA-RMENet, a U-Net-based deep learning model for path loss radio map estimation. The architecture combines an EfficientNetB5 encoder, attention-gated (AG) skip connections, and an Atrous Spatial Pyramid Pooling (ASPP) bottleneck, along with optional grid anchors and data augmentation. The authors evaluate two variants on the RadioMapSeer3D dataset and report a test RMSE of 0.0334 for EA-RMENet/DA with 0.022 seconds per sample, and an RMSE of 0.0406 on the ICASSP 2023 Radio Map Prediction Challenge late-submission leaderboard, where the model ranks third. The paper concludes that the model is an accurate and practical alternative to traditional path loss modelling, with potential for real-world RME.

Significance. If the reported results are robust, the paper makes a modest but useful incremental contribution: it integrates well-known components (EfficientNet encoder, AG skips, ASPP, data augmentation) in a U-Net for radio map estimation and demonstrates competitive accuracy and efficiency on a public benchmark and an external blind challenge. The third-place challenge result is a concrete, externally benchmarked point in favor of the architecture. However, the significance is limited by the absence of real-world validation, lack of code and appendices, and single-run metrics without error bars. The claimed "14.6% RMSE reduction from DA" and the encoder selection are not statistically substantiated. The central design is plausible, but the evidence as presented does not fully support the real-world deployment claims.

major comments (3)

[V-D, Table IV; Eq. (10)] All reported RMSE/MSE values (e.g., EA-RMENet/DA 0.0334 vs GA 0.0382 vs baseline 0.0391) are single-run point estimates. The claimed 14.6% DA improvement and 3.83% GA improvement are relative differences of order 0.001–0.006 in RMSE; without repeated seeds, error bars, or significance tests, the reader cannot tell whether these differences are real or due to training stochasticity. Please report mean±std over at least 3–5 independent runs for the baseline/GA/DA comparison and for Table IV, or provide a statistical test.
[III/Table I and VI/Table V] The conclusion states that EA-RMENet is "an accurate and practical alternative to traditional path loss modelling," and the abstract claims "potential for real-world RME." The only evaluations are RadioMapSeer ray-traced simulations and the ICASSP 2023 challenge, which is also a simulated dataset. The acknowledged +0.0072 RMSE distribution shift is simulation-to-simulation; no measured field data appear anywhere. To support the real-world claim, the authors should either validate on measurement data (e.g., drive-test or an established measurement-based benchmark) or explicitly restrict the claims to simulated urban radio maps in the title, abstract, and conclusion.
[IV-D, V-A, references [11][12]] The EfficientNetB5 encoder is selected from a 10-epoch ablation on RadioMapSeer2D using fixed hyperparameters, but the final model is trained for 40 epochs on RadioMapSeer3D. It is not established that the 10-epoch 2D ranking transfers to the final 3D task. Moreover, key implementation details (skip connection feature table D.7, decoder dimensions table D.8, ASPP diagram D.10, full ablation results table E.1, and the data pipeline redesign in §4.2.1) reside in appendices not included in the arXiv text, and no code or checkpoints are released. This prevents independent reimplementation and assessment of the reported numbers. Please provide the appendices as supplemental material and/or release code, and justify or re-run the encoder ablation on the 3D task.

minor comments (6)

[IV-B, Eq. (5)] The bracket notation in Eq. (5) is ambiguous. Please define it explicitly as clipping or clamping to [0,1] and clarify how the "cumulative slice encoding" differs from a one-hot height encoding.
[V-D, Table IV] RMSE and MSE in Table IV are not mutually consistent: 0.0382²=0.001459 vs. reported 0.001499, and 0.0334²=0.001116 vs. reported 0.001137. Please specify the rounding convention or correct the values.
[VI] Typo: "IASSP 2023 Radio Map Prediction Challenge" should be "ICASSP."
[Table V and References] Reference citations appear mismatched: Agile Method and PPNet both cite [16] (the PMNet challenge paper), and REM-Net+ is listed as [15] in Table V but [14] in the text/reference list. Please correct.
[IV-D] The text reports "14.73 million FLOPS". FLOPs are a count, not a rate, and the magnitude seems unusually low for an EfficientNetB5-based U-Net at 256×256 input. Please verify the unit and value.
[V-D, Figure 4] Figure 4 captions state the GA variant was plotted over 35 epochs, while Table III lists 40 training epochs. Please clarify the early stopping or checkpoint-selection procedure.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity — central RMSE figures come from held-out test data and an external blind challenge leaderboard.

full rationale

The paper's central quantitative claims are empirical evaluations, not derivations that reduce to their inputs. The RadioMapSeer3D result (RMSE 0.0334) is computed on a held-out test split defined in Table II (100 of 700 maps, 8,000 of 56,000 samples), with the model checkpoint selected by validation loss (Section V-B). The ICASSP 2023 challenge result (RMSE 0.0406) was obtained by blind submission to an external leaderboard (Section VI, Table V), so it is not fitted to the reported target. The preprocessing equations (1)-(5) are encodings of the input and ground-truth radio maps, not predictions defined in terms of the output. The architectural components (EfficientNet, AG skip connections, ASPP, data augmentation) are cited from external prior work [2][3][5][8][9], not from the authors' own unverified claims. The only self-references are to the authors' own thesis appendices [11][12] for implementation details and ablation plots; these are reproducibility gaps rather than load-bearing circularity, and they do not supply the central results or forbid alternatives. The limitation that all evaluation is on simulated radio maps is an external-validity concern, not a circularity concern. Overall, no prediction is forced by construction and no fitted parameter is renamed as a prediction.

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

No new physical quantity is proposed. The paper contributes a neural network architecture and an input representation; the performance claims rest on the trained weights, which are fitted to simulated data, and on manually chosen hyperparameters.

free parameters (6)

Building height slicing (N=12 slices) = 12
Height range 6.6-19.8 m divided into 12 cumulative slices (Eq. 4-5); chosen by hand with no ablation over N.
ASPP dilation rates = {2,3,4} plus 1x1 and global pooling
Dilation rates follow DeepLabV3 defaults; no RME-specific tuning or sensitivity study reported.
Training hyperparameters = lr 1e-3 to 1e-4 from epoch 15; weight decay 1e-5; batch 12; 40 epochs
Set manually; no grid search or sensitivity analysis.
Data augmentation policy = 4 rotations x horizontal flip (8 transforms)
Copied from PMNet [2]; no comparison of augmentation strengths or types.
Encoder scale = EfficientNetB5
Selected from a 10-epoch ablation on RadioMapSeer2D (Table E.1), not on the 3D target dataset and without repeated seeds.
Data split = 500/100/100 maps (40k/8k/8k samples)
Single fixed split; no cross-validation or multiple seeds.

assumptions (5)

domain assumption RadioMapSeer's ray-traced simulated radio maps are a valid proxy for real urban propagation.
All training/test/challenge data are simulations (Section III); no measured field data are used, so transfer to real networks is assumed.
domain assumption The 14/16 channel inputs (building footprint, Tx mask, 12 height slices, optional grid anchors) contain sufficient information to determine path loss.
The model is evaluated only on these features; material properties, terrain, vegetation, and weather are not included.
standard math Path loss values can be normalized linearly to [0,1] using the fixed range [-111,-75] dB and RMSE in normalized units is a meaningful accuracy measure.
This is a consistent affine transform; validity of the clipping and threshold choices is inherited from the RadioMapSeer dataset description [4].
domain assumption The fixed 500/100/100 map split gives an unbiased estimate of generalization to unseen maps.
No cross-validation or repeated splits; split quality and map diversity are assumed.
ad hoc to paper Choosing EfficientNetB5 via a 10-epoch ablation on RadioMapSeer2D transfers to the 3D task.
Encoder selection is performed on a different dataset and with shorter training than the final model (Section V-A, Appendix E).

how reviews work

0 comments

Cite this review

Pith. "Pith review of EA-RMENet -- Path Loss Prediction in Urban Environments using Deep Learning." pith.science (2026). https://pith.science/paper/P3UDAN2V

@misc{pith2026260716449,
  author       = {Pith},
  title        = {Pith review of: EA-RMENet -- Path Loss Prediction in Urban Environments using Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3UDAN2V}},
  note         = {Machine review of arXiv:2607.16449}
}

read the original abstract

Accurate path loss prediction is a critical component of wireless network planning. Current path loss prediction methods typically struggle to balance the trade-off between accuracy and computational efficiency. This paper proposes the Efficient Attention Radio Map Estimation Network (EA-RMENet) which is an image data-driven, deep learning (DL) model designed for radio map estimation (RME). EA-RMENet uses a U-Net framework with an EfficientNetB5 encoder, Attention Gated (AG) skip connections, and Atrous Spatial Pyramid Pooling (ASPP). The EfficientNet encoder uses compound scaling to balance accuracy and efficiency. AG skip connections suppress irrelevant features, and the ASPP captures a multi-scale context. The model has a test prediction RMSE of 0.0334 on the RadioMapSeer3D dataset with an inference time of 0.022 seconds/sample. In the ICASSP 2023 Radio Map Prediction Challenge, the model ranks third with a competitive RMSE of 0.0406 this highlights the models potential for real-world RME.

Figures

Figures reproduced from arXiv: 2607.16449 by the authors.

**Figure 3.** Attention gate data flow block diagram. F. Atrous Spatial Pyramid Pooling (ASPP) In CNNs, the receptive field defines the spatial range of the input influencing feature extraction [2]. For path loss modelling, relevant features span multiple scales: local effects [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** The EA-RMENet/GA variant exhibited signs of [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Both models achieved high accuracy and strong [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 2 linked inside Pith

[1]

RadioUNet: Fast Radio Map Estimation With Convolutional Neural Networks,

R. Levie and et al., “RadioUNet: Fast Radio Map Estimation With Convolutional Neural Networks,” IEEE Transactions on Wireless Communications, vol. 20, no. 6, pp. 4001 - 4015, 2021

2021
[2]

PMNet: Robust Pathloss Map Prediction via Supervised Learning,

J.-H. Lee and et al, “PMNet: Robust Pathloss Map Prediction via Supervised Learning,” in IEEE Global Communications Conference, Kuala Lumpur , 2023

2023
[3]

EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,

M. Tan and et al., “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” in Proceedings of the 36th International Conference on Machine Learning, PMLR, Long Beach, 2019

2019
[4]

Dataset of Pathloss and ToA Radio Maps with Localization Application,

C. Yapar and et al., “Dataset of Pathloss and ToA Radio Maps with Localization Application,” arXiv:2212.11777v4, 2022

arXiv 2022
[5]

Attention U-Net: Learning Where to Look for the Pancreas,

O. Oktay and et. al., “Attention U-Net: Learning Where to Look for the Pancreas,” arXiv:1804.03999v3, 2018

arXiv 2018
[6]

Transfer Learning and Double U-Net Empowered Wave Propagation Model in Complex Indoor Environments,

Z. Fu and et al., “Transfer Learning and Double U-Net Empowered Wave Propagation Model in Complex Indoor Environments,” IEEE Transactions on Antennas and Propagation, 2025

2025
[7]

Eff-UNet: A Novel Architecture for Semantic Segmentation in Unstructured Environment,

B. Baheti and et al., “Eff-UNet: A Novel Architecture for Semantic Segmentation in Unstructured Environment,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Seattle, 2020

2020
[8]

DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,

L.-C. Chen and et al, “DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 99, 2016

2016

Show all 16 references

[9]

Transformer based Radio Map Prediction Model for Dense Urban Environments,

Y. Tian and et al., “Transformer based Radio Map Prediction Model for Dense Urban Environments,” in 13th International Symposium on Antennas, Propagation and EM Theory (ISAPE), Zhuhai, 2021

2021
[10]

A Deep Learning Approach for Automatic Liver Segmentation: Attention U-Net with ASPP and CBAM Integration,

M. Diab and et al, “A Deep Learning Approach for Automatic Liver Segmentation: Attention U-Net with ASPP and CBAM Integration,” in 2025 6th International Conference on Bio-engineering for Smart Technologies (BioSMART), Paris, 2025

2025
[11]

Appendix D,

J. O'Shea, “Appendix D,” in Radio Map Prediction in Urban Environments, Dublin, DCU, 2025

2025
[12]

Appendix E,

J. O'Shea, “Appendix E,” in Radio Map Prediction in Urban Environments, Dublin, DCU, 2025

2025
[13]

Overview of the First Pathloss Radio Map Prediction Challenge,

C. Yapar and et al., “Overview of the First Pathloss Radio Map Prediction Challenge,” IEEE Open Journal of Signal Processing, vol. 5, pp. 948-963, 2024

2024
[14]

REM-Net+: 3D Radio Environment Map Construction Guided by Radio Propagation Model,

Q. Chen and et al., “REM-Net+: 3D Radio Environment Map Construction Guided by Radio Propagation Model,” TexhRix, 29 07 2024

2024
[15]

The First Pathloss Radio Map Prediction Challenge,

radiomapchallenge, “The First Pathloss Radio Map Prediction Challenge,” ICASSP, 2023. [Online]. Available: https://radiomapchallenge.github.io/results.html. [Accessed 10 07 2025]

2023
[16]

PMNet: Large-Scale Channel Prediction System for ICASSP 2023 First Pathloss Radio Map Prediction Challenge,

J.-H. Lee and et al., “PMNet: Large-Scale Channel Prediction System for ICASSP 2023 First Pathloss Radio Map Prediction Challenge,” in IEEE ICASSP 2023, Rhodes Island, 2023

2023

Pith tools

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

[1] [1]

RadioUNet: Fast Radio Map Estimation With Convolutional Neural Networks,

R. Levie and et al., “RadioUNet: Fast Radio Map Estimation With Convolutional Neural Networks,” IEEE Transactions on Wireless Communications, vol. 20, no. 6, pp. 4001 - 4015, 2021

2021

[2] [2]

PMNet: Robust Pathloss Map Prediction via Supervised Learning,

J.-H. Lee and et al, “PMNet: Robust Pathloss Map Prediction via Supervised Learning,” in IEEE Global Communications Conference, Kuala Lumpur , 2023

2023

[3] [3]

EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,

M. Tan and et al., “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” in Proceedings of the 36th International Conference on Machine Learning, PMLR, Long Beach, 2019

2019

[4] [4]

Dataset of Pathloss and ToA Radio Maps with Localization Application,

C. Yapar and et al., “Dataset of Pathloss and ToA Radio Maps with Localization Application,” arXiv:2212.11777v4, 2022

arXiv 2022

[5] [5]

Attention U-Net: Learning Where to Look for the Pancreas,

O. Oktay and et. al., “Attention U-Net: Learning Where to Look for the Pancreas,” arXiv:1804.03999v3, 2018

arXiv 2018

[6] [6]

Transfer Learning and Double U-Net Empowered Wave Propagation Model in Complex Indoor Environments,

Z. Fu and et al., “Transfer Learning and Double U-Net Empowered Wave Propagation Model in Complex Indoor Environments,” IEEE Transactions on Antennas and Propagation, 2025

2025

[7] [7]

Eff-UNet: A Novel Architecture for Semantic Segmentation in Unstructured Environment,

B. Baheti and et al., “Eff-UNet: A Novel Architecture for Semantic Segmentation in Unstructured Environment,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Seattle, 2020

2020

[8] [8]

DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,

L.-C. Chen and et al, “DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 99, 2016

2016