Pith. sign in

REVIEW 3 major objections 4 minor 25 references

Cement plants are detected more reliably by a short-wave infrared soil ratio (Band 7:6) than by thermal kiln heat, and a network with trainable Gabor filters, mixed pooling, and dilated convolutions reaches 90.6% accuracy.

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

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

T0 review · deepseek-v4-flash

2026-08-03 14:39 UTC pith:C6WCSL3I

load-bearing objection The paper's real finding is that a SWIR band ratio helps detect cement plants, but the FusionNet fusion claim is untested: no experiment uses all five inputs, and the reported tiny variance looks off. the 3 major comments →

arxiv 2512.19504 v2 pith:C6WCSL3I submitted 2025-12-22 cs.CV

FusionNet: Physics-Aware Representation Learning for Multi-Spectral and Thermal Data via Trainable Signal-Processing Priors

Georgios Voulgaris This is my paper
classification cs.CV
keywords cement plant detectionSWIR band ratiothermal infraredLandsat 8multi-spectral fusionGabor filterdilated convolutionremote sensing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the persistent environmental changes left by cement production—heated, dried, mineral-altered soil—are a stronger remote-sensing signal than the transient heat of the kiln itself. It proposes a spectral ratio, Landsat Band 7:6, to capture those soil alterations, and feeds it through a convolutional backbone whose first layer is a trainable Gabor filter bank, whose pooling mixes maximum and average activations, and whose final layers use dilated convolutions. Wrapped in an intermediate fusion network called FusionNet, this reaches 90.6% accuracy on the Band 7:6 dataset, outperforming standard CNN baselines and the strongest unimodal model. The paper also reports that ImageNet pretraining consistently hurts thermal-infrared performance, arguing that spectral priors matter more than generic RGB features. A sympathetic reader takes the core message to be that physics-aware band selection and signal-processing priors—not model scale—carry the performance.

Core claim

The central discovery is that indirect, persistent spectral signatures, specifically the Landsat 8 Band 7:6 short-wave infrared ratio, separate cement facilities from surrounding land cover more cleanly than direct thermal infrared emissions. The ablation results attribute the gain to each architectural component: the trainable Gabor-parameterised first convolution improves accuracy beyond mixed pooling, dilated receptive fields improve it further, and the full DGCNN backbone reaches 88.7% on the ratio. Embedding that backbone in FusionNet, a five-branch intermediate fusion architecture with channel attention, lifts accuracy to 90.6%. The paper presents this as validation that first-principl

What carries the argument

The load-bearing object is the Band 7:6 ratio: a per-pixel division of Landsat 8 band 7 (2.11–2.29 µm) by band 6 (1.57–1.65 µm), which compresses moisture and mineral signatures into one contrast channel and reduces atmospheric interference. The second mechanism is the DGCNN backbone: a first convolutional layer initialised with a bank of Gabor filters whose frequencies and orientations are updated by backpropagation, a per-layer mixture of max and average pooling with mixing weight decreasing from 1 to 0.2, and averaged dilated convolutions with rates 1, 3, 6, 9 before global pooling. In FusionNet these backbones run per spectral input; the outputs are concatenated and reweighted by a chann

Load-bearing premise

The argument depends on treating 1×1 km image chips as independent samples; if chips from the same cement plant appear in both the training and test splits, the 90.6% accuracy is partly memorization of known plants and would not transfer to unseen facilities.

What would settle it

Split the data by plant rather than by chip: hold out every chip belonging to a set of whole cement plants, train on the remaining plants, and measure accuracy on the held-out plants. If accuracy drops toward chance or the Band 7:6 advantage over TIR disappears, the central claim of generalisable soil-signal detection fails.

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

If this is right

  • If Band 7:6 is the discriminative signal, cement-plant monitoring can target persistent soil alteration rather than kiln temperature, making detection less vulnerable to background heat sources.
  • The DGCNN backbone transfers across both thermal and SWIR bands, so the architectural priors can be reused for other multi-spectral detection tasks.
  • ImageNet pretraining is counterproductive for thermal and short-wave infrared imagery; training from scratch aligns features with the sensing modality and yields better accuracy.
  • FusionNet's intermediate fusion with channel attention is a workable template for combining multiple spectral bands into a single detector.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper sets up a fully fused five-input experiment (bands 10, 11, 6, 7, and the 7:6 ratio together) but never reports its result; my inference is that the fusion claim remains open, and the decisive test is whether five-input fusion beats the 90.6% of Band 7:6 alone.
  • If the ratio truly tracks soil alteration from sustained heat, its discriminative power should increase with plant age or years of operation; checking accuracy against production start year would separate physical signal from spurious spectral correlation.
  • The reported standard deviation of ≤0.01% across five random splits is unusually low for roughly three thousand 1×1 km chips; an external replication with a plant-disjoint split would tell whether the margin over baselines is real.
  • The same Gabor-prior backbone could be probed on other persistent-pollution industries, such as steel or brick kilns, where heated soil and dust alter SWIR reflectance similarly.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes FusionNet, an intermediate multi-spectral fusion framework that combines Landsat 8 Thermal Infrared (Bands 10, 11) and Short Wave Infrared (Bands 6, 7, and a Band 7:6 ratio) through a backbone called DGCNN. DGCNN embeds a trainable Gabor-convolution layer, mixed max/avg pooling, and averaged dilated convolutions. The main claims are that the Band 7:6 SWIR ratio is more discriminative than direct thermal cues, that each architectural component contributes positively, that FusionNet outperforms SOTA baselines (peak 90.6% on Band 7:6), and that ImageNet pretraining degrades TIR performance. However, all reported experiments evaluate single spectral inputs separately; no experiment uses the full five-input fused representation, so the central fusion claim is not directly supported.

Significance. If the fusion result were actually demonstrated, the paper would contribute a practical physics-informed feature (Band 7:6 ratio) and a backbone design for industrial infrastructure mapping. The systematic ablation, per-class diagnostics (t-SNE, CAM, confusion matrices), and the negative ImageNet transfer-learning result are useful empirical observations. However, the current manuscript does not substantiate the titular 'multi-spectral data fusion' contribution: every table and figure reports accuracy for individual bands or the ratio, not for a fused multi-band input. The implausibly low variance across splits also raises concerns about the reliability of the reported numbers. The paper's strengths are its real-world dataset and the breadth of architectural ablations, but these are undermined by the missing fusion experiment and unresolved evaluation methodology.

major comments (3)
  1. [Section II-B, Tables II-III] The central claim is that FusionNet integrates TIR and SWIR inputs, yet no experiment with all five inputs fused together is reported. Section II-B states the model 'was evaluated across TIR (Bands 10 and 11), SWIR (Bands 6 and 7), a geological SWIR ratio (Band 7:6), and a fused dataset comprising all five inputs.' Tables II and III and Figures 7-8 list only the five individual spectral inputs. The maximum accuracy of 90.6% appears in Table III under 'FusionNet DGCNN' in the Band 7:6 column, which is a single-input result, not a fused result. The paper's conclusion that FusionNet 'jointly leverages TIR and SWIR information' is therefore untested. Please add results for the full five-channel fused input, and clarify how the model is configured when fewer than five inputs are used (e.g., are remaining branches zeroed or omitted?).
  2. [Section III-A, standard deviation statement] The statement 'standard deviations were consistently <=0.01% across all experiments' is statistically implausible for classification accuracy on test sets of the sizes implied by Table I. For a binary classification problem with roughly 500-600 test samples, the binomial standard error of accuracy is on the order of 1-2 percentage points. A standard deviation below 0.01% across five independent random splits would require near-identical test sets or a miscalculated metric. Please report the per-split accuracies, exact test-set sizes, and confidence intervals. If the splits are not independent (e.g., chips from the same cement plant appear in both training and test), spatial autocorrelation would inflate accuracy and invalidate the comparison.
  3. [Section II-A / III-C, Band 7:6 feature selection] The Band 7:6 ratio is presented as a physics-informed feature and then evaluated on the same dataset used to motivate it. If the ratio was selected after inspecting performance on this dataset, the reported advantage over individual bands may reflect selection bias. Please clarify whether any feature or model selection used the test data, and ideally provide an independent validation set or a nested cross-validation procedure. At present, the claim that the SWIR ratio is more discriminative than TIR is supported only by in-sample comparisons.
minor comments (4)
  1. [Abstract / Section III-C] The abstract reports FusionNet reaching 90.6% without specifying that this is for the Band 7:6 input alone. The contribution list in Section I similarly says 'achieving state-of-the-art performance with 90.6% accuracy' with no qualifier. Please disambiguate single-band versus fused results throughout.
  2. [Section II-C / Figure 7] Section II-C mentions a 'fused dataset comprising all five inputs' and Figure 7 caption says 'Across TIR (Bands 11–10) and SWIR (Bands 7–6 and Ratio 7:6) Inputs,' which are inconsistent. If the five-input fusion experiment is added, specify the fusion protocol (e.g., channel-wise concatenation before the backbone) and include it in the figure captions.
  3. [Section II-C] The training section does not state how the 80/20 split is performed at the chip level. Were chips from the same cement plant allowed to appear in both training and test? This is important for assessing spatial independence and should be explicitly described.
  4. [General] No code or data availability statement is provided. Given the difficulty of reproducing remote-sensing deep learning results, please include a statement or a link to the code/data.

Circularity Check

0 steps flagged

No circular reduction found; the classification benchmark is self-contained. Main issues are a minor non-load-bearing self-citation for the DGCNN backbone and a missing all-five-input fusion experiment (a correctness concern, not circularity).

full rationale

The paper's central numbers (88.7%, 90.6%) are measured accuracies from trained CNNs on held-out splits of Landsat 8 chips; they are not obtained by fitting an equation to the target and then renaming the fit a prediction. The Band 7:6 ratio is a fixed spectral ratio motivated by cited soil-moisture/reflectance literature (ref. [15]) and by physical reasoning in Section II-A; it is not fitted to the labels, so the claim that it is discriminative is an empirical benchmark result rather than a self-definitional reduction. The DGCNN backbone is imported from the authors' prior work [16] ('These backbones incorporate ... (Voulgaris et al. [16])', Section II-B), but the contribution of each component is independently measured in Table II on the current dataset, so the self-citation is not load-bearing. No uniqueness theorem or ansatz is smuggled in via citation; the Gabor, MixPool and dilated-convolution priors are stated explicitly and ablated. The paper does contain an omitted-support problem: Section II-B says the model 'was evaluated across ... a fused dataset comprising all five inputs,' yet Tables II-III and Figs. 3, 5, 7, 8 report only the five individual spectral inputs, and the claimed 90.6% is the Band 7:6 (single-input) column of Table III. That is a missing experiment relative to the fusion claim, not a circular step. Overall score 2 reflects the minor non-load-bearing self-citation; the derivation itself is not circular.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on several hand-chosen hyperparameters and domain assumptions from remote sensing. No new entities are introduced. The most fragile background assumptions are spatial independence of the data split and the credibility of the reported near-zero variance.

free parameters (4)
  • mixed pooling weight schedule α_l = 1.0 decreasing by 0.2 per layer (hand-set)
    Manual schedule for the trade-off between max and average pooling across layers; no tuning or ablation reported.
  • dilated convolution rates = 1, 3, 6, 9
    Hand-chosen dilation rates to increase receptive field; no search described.
  • class weight = [3, 1]
    Hand-chosen to balance cement vs landcover classes 4:1; no sensitivity analysis.
  • Gabor parameter grid (ω_n, θ_m) = 5 frequencies, 8 orientations
    Initialization grid for the trainable Gabor layer; chosen by hand.
axioms (4)
  • domain assumption Band 7:6 ratio of Landsat 8 is sensitive to soil moisture and mineral composition changes caused by heat emissions
    Assumed from remote-sensing literature (reference [15]) and used to justify feature selection; not independently verified in this dataset.
  • domain assumption Global Database of Cement Production Assets provides accurate, complete locations of active cement plants in China
    Used to create the labels; errors in the database propagate into the classification target.
  • ad hoc to paper Randomly splitting image chips yields train/test sets with independent distributions
    Unstated in Section II-C; if spatial autocorrelation exists between chips of the same plant, this is violated.
  • domain assumption The reported standard deviations (≤0.01%) are correct
    Implied by the text; implausible for a small test set with 5 random splits.

pith-pipeline@v1.3.0-alltime-deepseek · 11410 in / 13839 out tokens · 122178 ms · 2026-08-03T14:39:14.345023+00:00 · methodology

0 comments
read the original abstract

Cement production underpins global infrastructure but contributes approximately 7% of anthropogenic CO2 emissions, making accurate monitoring of production facilities essential for sustainable development. Existing remote sensing approaches rely predominantly on thermal signatures from kiln operations, which can be confounded by background heat sources and fail to capture persistent environmental alterations. This study introduces a physics-informed methodology that exploits multi-spectral features, particularly a geological Short Wave Infrared (SWIR) ratio, to detect soil property changes induced by sustained industrial heat emissions. This work proposes FusionNet, an intermediate multi-spectral data fusion framework that integrates Thermal Infrared (TIR) and SWIR inputs through a specialised backbone, embedding differential signal processing priors within a convolutional layer, mixed pooling, and wider receptive field. Systematic ablation studies confirm that each architectural component contributes to performance gains, with DGCNN achieving a 4.1-6.8% accuracy improvement over conventional CNNs. On the SWIR ratio dataset, FusionNet attains a maximum of 90.6%, outperforming state-of-the-art baselines across five spectral configurations and exceeding the strongest unimodal model by 1.1%. Transfer learning experiments reveal that ImageNet pretraining degrades TIR and SWIR performance, underscoring the importance of modality-aware training for cross-spectral applications. Overall, the results demonstrate that combining physics-aware feature selection with principled deep learning architectures enables robust, high-accuracy detection of cement production facilities, offering a reliable framework for industrial infrastructure monitoring

Figures

Figures reproduced from arXiv: 2512.19504 by Georgios Voulgaris.

Figure 1
Figure 1. Figure 1: a) Cement Chip and Surrounding Landcover, b) Each Image is Comprised of 3-Channels, Each Channel Represents a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: a) FusionNet: TIR (Bands 11-10) and SWIR (Bands 7-6 and Ratio 7:6), b) Backbone Networks: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: CNN5 vs DGCNN Average Accuracy (%) Scores Across TIR (Bands 11–10) and SWIR (Bands 7–6 and Ratio 7:6) Inputs, Evaluated per Cement and Landcover Classes. The Proposed DGCNN Outperforms the Conventional Convolu￾tional Backbone in Both Thermal and SWIR Datasets, with the Proposed Ratio Yielding the Highest Overall Accuracy [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: t-Distributed Stochastic Neighbour Embedding (t-SNE) a) CNN5, b) DGCNN per Cement and Landcover Classes. The [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Confusion Matrix Results a) CNN5, b) DGCNN per [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Class Activation Maps, CNN5 vs DGCNN Feature [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FusionNet Average Accuracy (%) Scores Across TIR (Bands 11–10) and SWIR (Bands 7–6 and Ratio 7:6) Inputs, Evaluated per Cement and Landcover Classes. The Proposed FusionNet Outperforms all State-Of-The-Art Models in Both Thermal and SWIR Datasets, with The Geological SWIR Ratio Yielding the Highest Overall Accuracy. In contrast, the DGCNN model demonstrates a more fo￾cused and detailed activation pattern, … view at source ↗
Figure 8
Figure 8. Figure 8: FusionNet Recall Scores: TIR (Bands 11-10) & SWIR [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of SOTA CNN models (Accuracy ( [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

25 extracted references · 3 linked inside Pith

  1. [1]

    Historical trend and decarbonization pathway of china’s cement industry: A literature review,

    Y . Wang, H. Yi, X. Tang, Y . Wang, H. An, and J. Liu, “Historical trend and decarbonization pathway of china’s cement industry: A literature review,”Science of The Total Environment, vol. 891, p. 164580, 2023

    2023

  2. [2]

    Emission reduction of greenhouse gases from the cement industry,

    C. A. Hendriks, E. Worrell, D. De Jager, K. Blok, and P. Riemer, “Emission reduction of greenhouse gases from the cement industry,” inProceedings of the fourth international conference on greenhouse gas control technologies, vol. 30. IEA GHG R&D Programme Interlaken Austria, 1998, pp. 939–944

    1998

  3. [3]

    Anal- ysis on differences of carbon dioxide emission from cement production and their major determinants,

    T. Gao, L. Shen, M. Shen, F. Chen, L. Liu, and L. Gao, “Anal- ysis on differences of carbon dioxide emission from cement production and their major determinants,”Journal of Cleaner Production, vol. 103, pp. 160–170, 2015

    2015

  4. [4]

    Korczak, M

    K. Korczak, M. Kocha ´nski, and T. Skoczkowski, “Mitigation options for decarbonization of the non-metallic minerals indus- try and their impacts on costs, energy consumption and ghg emissions in the eu-systematic literature review,”Journal of Cleaner Production, vol. 358, p. 132006, 2022

    2022

  5. [5]

    Asharfi, Y

    A. Asharfi, Y . Erfanifard, F. Amiraslani, A. D. Boloorani, and A. Mousivand, “Quantitative assessment of forest ecosystem stress caused by cement plant pollution using in situ measure- ments and sentinel-2 satellite data in a part of the unesco world heritage site,”European Journal of Environmental Sciences, vol. 10, no. 1, pp. 22–31, 2020

    2020

  6. [6]

    Global database of cement production assets and upstream suppliers,

    N. Tkachenko, K. Tang, M. McCarten, S. Reece, D. Kampmann, C. Hickey, M. Bayaraa, P. Foster, C. Layman, C. Rossiet al., “Global database of cement production assets and upstream suppliers,”Scientific Data, vol. 10, no. 1, p. 696, 2023

    2023

  7. [7]

    Extending satellite remote sensing to local scales: land and water resource monitoring using high-resolution imagery,

    K. E. Sawaya, L. G. Olmanson, N. J. Heinert, P. L. Brezonik, and M. E. Bauer, “Extending satellite remote sensing to local scales: land and water resource monitoring using high-resolution imagery,”Remote sensing of Environment, vol. 88, no. 1-2, pp. 144–156, 2003

    2003

  8. [8]

    Carbon dioxide emissions from the global cement industry,

    E. Worrell, L. Price, N. Martin, C. Hendriks, and L. O. Meida, “Carbon dioxide emissions from the global cement industry,” Annual review of energy and the environment, vol. 26, no. 1, pp. 303–329, 2001

    2001

  9. [9]

    Low carbon measures for cement plant–a review,

    S. A. Ishak and H. Hashim, “Low carbon measures for cement plant–a review,”Journal of Cleaner Production, vol. 103, pp. 260–274, 2015

    2015

  10. [10]

    Emission reduction of greenhouse gases from the cement in- dustry, greenhouse gas control technologies conference,

    C. Hendriks, E. Worrell, D. de Jager, K. Blok, and P. Riemer, “Emission reduction of greenhouse gases from the cement in- dustry, greenhouse gas control technologies conference,” 2004

    2004

  11. [11]

    De- tection and characterisation of pollutant assets with ai and eo to prioritise green investments: the geoasset framework,

    C. Rossi, N. Tkachenko, M. Bayaraa, P. Foster, S. Reece, K. Scott, G. V oulgaris, C. Christiaen, and M. McCarten, “De- tection and characterisation of pollutant assets with ai and eo to prioritise green investments: the geoasset framework,” in IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2022, pp. 7717–7720

    2022

  12. [12]

    Identifying industrial heat sources using time-series of the viirs nightfire product with an object-oriented approach,

    Y . Liu, C. Hu, W. Zhan, C. Sun, B. Murch, and L. Ma, “Identifying industrial heat sources using time-series of the viirs nightfire product with an object-oriented approach,”Remote Sensing of Environment, vol. 204, pp. 347–365, 2018

    2018

  13. [13]

    Assessing heavy industrial heat source distribution in china using real-time viirs active fire/hotspot data,

    C. Ma, J. Yang, F. Chen, Y . Ma, J. Liu, X. Li, J. Duan, and R. Guo, “Assessing heavy industrial heat source distribution in china using real-time viirs active fire/hotspot data,”Sustainabil- ity, vol. 10, no. 12, p. 4419, 2018

    2018

  14. [14]

    An approach to large-scale cement plant detection using multisource remote sensing imagery,

    T. Li, C. Ma, Y . Lv, R. Liao, J. Yang, and J. Liu, “An approach to large-scale cement plant detection using multisource remote sensing imagery,”Remote Sensing, vol. 16, no. 4, p. 729, 2024

    2024

  15. [15]

    Effective band ratio of landsat 8 images based on vnir-swir reflectance spectra of topsoils for soil moisture mapping in a tropical region,

    D. Ngo Thi, N. T. T. Ha, Q. Tran Dang, K. Koike, and N. Mai Trong, “Effective band ratio of landsat 8 images based on vnir-swir reflectance spectra of topsoils for soil moisture mapping in a tropical region,”Remote Sensing, vol. 11, no. 6, p. 716, 2019

    2019

  16. [16]

    Seasonal domain shift in the global south: dataset and deep features analysis,

    G. V oulgaris, A. Philippides, J. Dolley, J. Reffin, F. Mar- shall, and N. Quadrianto, “Seasonal domain shift in the global south: dataset and deep features analysis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2116–2124

    2023

  17. [17]

    Network in network,

    M. Lin, Q. Chen, and S. Yan, “Network in network,”arXiv preprint arXiv:1312.4400, 2013

    Pith/arXiv arXiv 2013

  18. [18]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  19. [19]

    Semantic image segmentation with deep convolutional nets and fully connected crfs,

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, “Semantic image segmentation with deep convolutional nets and fully connected crfs,”arXiv preprint arXiv:1412.7062, 2014

    Pith/arXiv arXiv 2014

  20. [20]

    Deeplab: Semantic image segmentation with deep con- volutional nets, atrous convolution, and fully connected crfs,

    ——, “Deeplab: Semantic image segmentation with deep con- volutional nets, atrous convolution, and fully connected crfs,” IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 4, pp. 834–848, 2017

    2017

  21. [21]

    Re- visiting dilated convolution: A simple approach for weakly-and semi-supervised semantic segmentation,

    Y . Wei, H. Xiao, H. Shi, Z. Jie, J. Feng, and T. S. Huang, “Re- visiting dilated convolution: A simple approach for weakly-and semi-supervised semantic segmentation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7268–7277

    2018

  22. [22]

    Stochastic neighbor embedding,

    G. E. Hinton and S. Roweis, “Stochastic neighbor embedding,” Advances in neural information processing systems, vol. 15, 2002

    2002

  23. [23]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2921–2929

    2016

  24. [24]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan, “Very deep convolutional networks for large-scale image recognition,”arXiv preprint arXiv:1409.1556, 2014

    Pith/arXiv arXiv 2014

  25. [25]

    Efficientnet: Rethinking model scaling for convolutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 6105–6114

    2019