pith. sign in

arxiv: 2605.21352 · v1 · pith:DGQCXH4Ynew · submitted 2026-05-20 · 💻 cs.LG · cs.CE· cs.ET

Classification of Single and Mixed Partial Discharges under Switching Voltage Using an AWA-CNN Framework

Md Rafid Kaysar Shagor , Zannatul Ferdousy Mouri , Farhina Haque , Anindya Bijoy Das This is my paper

Pith reviewed 2026-05-21 05:25 UTC · model grok-4.3

classification 💻 cs.LG cs.CEcs.ET
keywords partial discharge classificationswitching voltageAWA patternconvolutional neural networkpower electronicsmixed discharge sourcesfault diagnosis
0
0 comments X

The pith

AWA visual patterns enable CNNs to classify six single and mixed partial discharge sources under switching voltage at over 96 percent accuracy.

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

The paper develops an Amplitude-Width-Area pattern that turns time-domain partial discharge pulses into color-coded images for use with convolutional neural networks. These patterns are tested on six conditions: corona, internal, surface, and the three pairwise mixtures, under switching voltage where discharges cluster at transitions. CNN models reach testing accuracy above 96 percent while a Random Forest baseline reaches only 73.33 percent. The results matter because fast-switching power electronics are now common and reliable source identification helps prevent insulation damage. The work shows that the visual encoding captures source-specific pulse features well enough for multi-class distinction.

Core claim

The central claim is that partial discharge pulses under switching voltage can be mapped to AWA patterns with amplitude and area as coordinate axes and width encoded by color, producing distinguishable distributions for corona, internal, surface, corona+internal, corona+surface, and internal+surface conditions. When these patterns serve as input to CNN architectures such as InceptionV3 and ResNet-18, classification accuracy exceeds 96 percent on held-out test data, substantially above the 73.33 percent achieved by Random Forest on the same representation.

What carries the argument

The AWA pattern is a two-dimensional visual map generated from each PD pulse, with amplitude and area as axes and width represented by color, that converts raw time-series signals into images carrying source-dependent structure for CNN input.

If this is right

  • AWA patterns exhibit source-dependent distributions that separate the six single and mixed PD conditions visually.
  • CNN classifiers applied to AWA patterns outperform Random Forest by a wide margin on the same data.
  • The representation works under switching-voltage excitation where activity concentrates at voltage transitions.
  • Both single-source and mixed-source PD conditions can be handled within the same six-class framework.

Where Pith is reading between the lines

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

  • The same pulse-to-image encoding could be applied to other non-sinusoidal voltage waveforms encountered in modern converters.
  • Embedding AWA-CNN pipelines in online monitoring hardware might allow continuous source tracking without manual feature engineering.
  • Validation on field-collected data from actual power-electronics installations would test whether laboratory patterns generalize.
  • Color-encoded feature maps of this type may transfer to other pulse-classification problems in high-voltage diagnostics.

Load-bearing premise

The laboratory-collected partial discharge data under switching voltage conditions is representative of field behavior and the AWA mapping retains enough source-specific information to support reliable multi-class separation.

What would settle it

A new collection of switching-voltage PD experiments that produces AWA patterns with large overlap between the six source classes, driving CNN test accuracy below 80 percent.

Figures

Figures reproduced from arXiv: 2605.21352 by Anindya Bijoy Das, Farhina Haque, Md Rafid Kaysar Shagor, Zannatul Ferdousy Mouri.

Figure 1
Figure 1. Figure 1: Fabricated PD test samples and verification samples for (a) surface PD, (b) internal PD, (c) mixed PD, (d) corona PD, (e) ceramic-plate surface-PD verification sample, and (f) cured epoxy internal-PD verification sample. Numerous studies have been conducted to investigate the PD behavior under PWM and pulse-type excitation. Under repetitive PWM pulses, Guo et al. reported that PD activity is mainly associa… view at source ↗
read the original abstract

The growing use of fast-switching power electronics has made partial discharge (PD) analysis under switching-voltage excitation increasingly important, yet more challenging than under sinusoidal conditions due to activity concentrated at voltage transitions. This work presents an Amplitude-Width-Area (AWA) pattern representation for source-oriented PD analysis under switching-voltage excitation. In the proposed method, time domain PD pulses are characterized using pulse amplitude, width, and area, and mapped into a visual pattern where amplitude and area define the coordinate axes and width is encoded by color. The generated AWA patterns are used to distinguish six single and mixed PD source conditions: corona, internal, surface, corona+internal, corona+surface, and internal+surface. To evaluate the classification capability of the proposed representation, a Random Forest baseline and two Convolutional Neural Network (CNN) models, InceptionV3 and ResNet-18, are compared. The AWA patterns show distinguishable source-dependent distributions, and CNN-based classification achieves testing accuracy above 96%, compared with 73.33% for Random Forest. The results indicate that AWA patterns provide a visual representation of PD pulses suitable for multi-class PD source classification under switching-voltage excitation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper proposes an Amplitude-Width-Area (AWA) visual pattern representation for partial discharge (PD) pulses under switching-voltage excitation. Time-domain pulses are mapped to 2D patterns with amplitude and area as axes and width encoded by color; these patterns are used to classify six single and mixed PD sources (corona, internal, surface, and their combinations). CNN models (ResNet-18 and InceptionV3) are compared against a Random Forest baseline, with the claim that AWA patterns exhibit source-dependent distributions and yield CNN test accuracy above 96% versus 73.33% for RF.

Significance. If the reported accuracy is supported by adequate independent samples and rigorous validation, the AWA representation could provide a practical visual encoding for multi-class PD source identification in the challenging regime of switching-voltage excitation, where activity is concentrated at transitions. The explicit baseline comparison and use of standard CNN architectures are positive aspects of the evaluation.

major comments (1)
  1. [Results / Experimental Setup] Results section (and associated experimental description): The headline claim of CNN test accuracy above 96% on the six-class AWA patterns is presented without any information on total pulse count, per-class sample sizes, class balance, or the data-splitting protocol (e.g., whether test patterns derive from held-out experimental sessions or random splits within the same runs). Under switching-voltage conditions, where PD events are sparse and temporally clustered, this omission leaves the performance gap over the 73.33% Random Forest baseline vulnerable to explanations such as overfitting or data leakage rather than the discriminative power of the AWA representation.
minor comments (1)
  1. [Abstract] Abstract: The statement 'testing accuracy above 96%' is given without the exact per-model figures, standard deviations, or any mention of the number of trials.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful review and constructive comment. We agree that additional details on the dataset and validation protocol are required to strengthen the manuscript and will incorporate them in the revision.

read point-by-point responses

  1. Referee: [Results / Experimental Setup] Results section (and associated experimental description): The headline claim of CNN test accuracy above 96% on the six-class AWA patterns is presented without any information on total pulse count, per-class sample sizes, class balance, or the data-splitting protocol (e.g., whether test patterns derive from held-out experimental sessions or random splits within the same runs). Under switching-voltage conditions, where PD events are sparse and temporally clustered, this omission leaves the performance gap over the 73.33% Random Forest baseline vulnerable to explanations such as overfitting or data leakage rather than the discriminative power of the AWA representation.

    Authors: We agree that the omission of these details is a limitation in the current manuscript and leaves the results open to the interpretations noted. In the revised version we will add a dedicated subsection (or expanded paragraph) under Experimental Setup that reports: (i) the total number of extracted PD pulses, (ii) the exact per-class sample counts and any resulting class imbalance, and (iii) the data-splitting protocol, explicitly stating that test patterns are taken from completely held-out experimental sessions rather than random intra-run splits. These additions will allow readers to assess independence and will support the claim that the observed accuracy gap reflects the discriminative properties of the AWA representation. revision: yes

Circularity Check

0 steps flagged

No circularity; standard ML classification on transformed features with held-out evaluation

full rationale

The paper defines an AWA (Amplitude-Width-Area) visual pattern transformation from raw PD pulses and feeds the resulting images into off-the-shelf CNNs (ResNet-18, InceptionV3) for 6-class source classification. Reported accuracies (>96% CNN vs 73.33% Random Forest) are obtained via conventional supervised training and testing on held-out data; no equations, parameters, or uniqueness claims are fitted to the target metric and then re-presented as predictions. No self-citations, ansatzes, or renamings of prior results appear in the load-bearing steps. The derivation chain is therefore self-contained empirical evaluation rather than a reduction to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim depends on the validity of the AWA pattern as a discriminative representation and the generalization ability of the CNN models trained on the specific dataset.

axioms (1)
  • domain assumption PD pulses under switching voltage can be characterized by amplitude, width, and area in a way that their visual mapping distinguishes sources.
    Invoked in the method description for creating AWA patterns.

pith-pipeline@v0.9.0 · 5765 in / 1263 out tokens · 53992 ms · 2026-05-21T05:25:18.962246+00:00 · methodology

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

34 extracted references · 34 canonical work pages

  1. [1]

    Power Electronics Based on Wide -Bandgap Semiconductors: Opportunities and Challenges,

    G. Iannaccone, C. Sbrana, I. Morelli, and S. Strangio, “Power Electronics Based on Wide -Bandgap Semiconductors: Opportunities and Challenges,” IEEE Access , vol. 9, pp. 139446 –139456, 2021, doi: 10.1109/ACCESS.2021.3118897

    work page doi:10.1109/access.2021.3118897 2021

  2. [2]

    Recent Advancements in Wide Band Semiconductors (SiC and GaN) Technology for Future Devices,

    S. Singh, T. Chaudhary, and G. Khanna, “Recent Advancements in Wide Band Semiconductors (SiC and GaN) Technology for Future Devices,” Silicon, vol. 14, no. 11, pp. 5793 –5800, Jul. 2022, doi: 10.1007/s12633 -021- 01362-3

    work page doi:10.1007/s12633 2022

  3. [3]

    Wide bandgap (WBG) semiconductor power converters for DC microgrid applications,

    K. Shenai, “Wide bandgap (WBG) semiconductor power converters for DC microgrid applications,” in 2015 IEEE First International Conference on DC Microgrids (ICDCM) , Atlanta, GA, USA: IEEE, Jun. 2015, pp. 263 –

    work page 2015

  4. [4]

    doi: 10.1109/ICDCM.2015.7152051

    work page doi:10.1109/icdcm.2015.7152051 2015

  5. [5]

    A Review of Power Electronic Converters for Electric Aircrafts,

    S. M. S. H. Rafin, M. A. Haque, R. Islam, and O. A. Mohammed, “A Review of Power Electronic Converters for Electric Aircrafts,” in 2023 Fourth International Symposium on 3D Power Electronics Integration and Manufacturing (3D-PEIM), Miami, FL, USA: IEEE, Feb. 2023, pp. 1 –8. doi: 10.1109/3D-PEIM55914.2023.10052535

    work page doi:10.1109/3d-peim55914.2023.10052535 2023

  6. [6]

    Wide -Bandgap Power Semiconductors for Electric Vehicle Systems: Challenges and Trends,

    T. Van Do, J. P. F. Trovao, K. Li, and L. Boulon, “Wide -Bandgap Power Semiconductors for Electric Vehicle Systems: Challenges and Trends,” IEEE Veh. Technol. Mag. , vol. 16, no. 4, pp. 89 –98, Dec. 2021, doi: 10.1109/MVT.2021.3112943

    work page doi:10.1109/mvt.2021.3112943 2021

  7. [7]

    Advances in Power Conversion and Drives for Shipboard Systems,

    F. Wang, Z. Zhang, T. Ericsen, R. Raju, R. Burgos, and D. Boroyevich, “Advances in Power Conversion and Drives for Shipboard Systems,” Proc. IEEE , vol. 103, no. 12, pp. 2285 –2311, Dec. 2015, doi: 10.1109/JPROC.2015.2495331

    work page doi:10.1109/jproc.2015.2495331 2015

  8. [8]

    A Comprehensive Review of Mitigation Strategies to Address Insulation Challenges Within High -Voltage, High-Power-Density (U)WBG Power Module Packages,

    P. Adhikari and M. Ghassemi, “A Comprehensive Review of Mitigation Strategies to Address Insulation Challenges Within High -Voltage, High-Power-Density (U)WBG Power Module Packages,” IEEE Trans. Dielectr. Electr. Insul. , vol. 31, no. 5, pp. 2676 –2700, Oct. 2024, doi: 10.1109/TDEI.2024.3382070

    work page doi:10.1109/tdei.2024.3382070 2024

  9. [9]

    Mitigating PWM Voltage -Induced Partial Discharge by Electrets,

    F. Haque and C. Park, “Mitigating PWM Voltage -Induced Partial Discharge by Electrets,” IEEE Trans. Dielectr. Electr. Insul., vol. 31, no. 2, pp. 763–771, Apr. 2024, doi: 10.1109/TDEI.2024.3363123

    work page doi:10.1109/tdei.2024.3363123 2024

  10. [10]

    Overview and Partial Discharge Analysis of Power Transformers: A Literature Review,

    M. R. Hussain, S. S. Refaat, and H. Abu-Rub, “Overview and Partial Discharge Analysis of Power Transformers: A Literature Review,” IEEE Access, vol. 9, pp. 64587–64605, 2021, doi: 10.1109/ACCESS.2021.3075288

    work page doi:10.1109/access.2021.3075288 2021

  11. [11]

    Detection, Classification and Location of Sources of Partial Discharges Using the Radiometric Method: Trends, Challenges and Open Issues,

    G. V. R. Xavier, H. S. Silva, E. G. Da Costa, A. J. R. Serres, N. B. Carvalho, and A. S. R. Oliveira, “Detection, Classification and Location of Sources of Partial Discharges Using the Radiometric Method: Trends, Challenges and Open Issues,” IEEE Access, vol. 9, pp. 110787–110810, 2021, doi: 10.1109/ACCESS.2021.3102888

    work page doi:10.1109/access.2021.3102888 2021

  12. [12]

    A Review of Knowledge -Based Defect Identification via PRPD Patterns in High Voltage Apparatus,

    T. Shahsavarian et al. , “A Review of Knowledge -Based Defect Identification via PRPD Patterns in High Voltage Apparatus,” IEEE Access, vol. 9, pp. 77705–77728, 2021, doi: 10.1109/ACCESS.2021.3082858

    work page doi:10.1109/access.2021.3082858 2021

  13. [13]

    Partial discharge signal interpretation for generator diagnostics,

    C. Hudon and M. Belec, “Partial discharge signal interpretation for generator diagnostics,” IEEE Trans. Dielectr. Electr. Insul., vol. 12, no. 2, pp. 297–319, Apr. 2005, doi: 10.1109/TDEI.2005.1430399

    work page doi:10.1109/tdei.2005.1430399 2005

  14. [14]

    The different stages of PRPD pattern for positive point to plane corona driven by a DC voltage containing ripple,

    T. Dezenzo, T. Betz, and A. Schwarzbacher, “The different stages of PRPD pattern for positive point to plane corona driven by a DC voltage containing ripple,” IEEE Trans. Dielectr. Electr. Insul., vol. 25, no. 1, pp. 30– 37, Feb. 2018, doi: 10.1109/TDEI.2018.006670

    work page doi:10.1109/tdei.2018.006670 2018

  15. [15]

    Convolution Neural Network in Partial Discharge Source Identification: Performance Analysis for Lab-Based Data and Electric Machine,

    M. R. K. Shagor, Z. F. Mouri, and F. Haque, “Convolution Neural Network in Partial Discharge Source Identification: Performance Analysis for Lab-Based Data and Electric Machine,” in 2025 IEEE Electrical Insulation Conference (EIC), South Padre Island, TX, USA: IEEE, Jun. 2025, pp. 1 –5. doi: 10.1109/EIC63069.2025.11123206

    work page doi:10.1109/eic63069.2025.11123206 2025

  16. [16]

    Characterization of Partial Discharges in High-Frequency Transformer Under PWM Pulses,

    Z. Guo, A. Q. Huang, R. E. Hebner, G. C. Montanari, and X. Feng, “Characterization of Partial Discharges in High-Frequency Transformer Under PWM Pulses,” IEEE Trans. Power Electron., vol. 37, no. 9, pp. 11199–11208, Sep. 2022, doi: 10.1109/TPEL.2022.3169747

    work page doi:10.1109/tpel.2022.3169747 2022

  17. [17]

    A Partial Discharge Classification Approach in Shipboard Power System Under Power Electronics Switching Voltage,

    M. R. K. Shagor and F. Haque, “A Partial Discharge Classification Approach in Shipboard Power System Under Power Electronics Switching Voltage,” in 2025 IEEE Electric Ship Technologies Symposium (ESTS) , Alexandria, VA, USA: IEEE, Aug. 2025, pp. 223 –229. doi: 10.1109/ESTS62818.2025.11152453

    work page doi:10.1109/ests62818.2025.11152453 2025

  18. [18]

    Partial Discharge Features for Power Electronic Transformers Under High-Frequency Pulse Voltage,

    J. Jiang, B. Zhang, Z. Li, P. Ranjan, J. Chen, and C. Zhang, “Partial Discharge Features for Power Electronic Transformers Under High-Frequency Pulse Voltage,” IEEE Trans. Plasma Sci. , vol. 49, no. 2, pp. 845 –853, Feb. 2021, doi: 10.1109/TPS.2021.3053960

    work page doi:10.1109/tps.2021.3053960 2021

  19. [19]

    Classification of Partial Discharges Originating From Multilevel PWM Using Machine Learning,

    E. Balouji, T. Hammarstrom, and T. McKelvey, “Classification of Partial Discharges Originating From Multilevel PWM Using Machine Learning,” IEEE Trans. Dielectr. Electr. Insul. , vol. 29, no. 1, pp. 287 –294, Feb. 2022, doi: 10.1109/TDEI.2022.3148461

    work page doi:10.1109/tdei.2022.3148461 2022

  20. [20]

    IEC 60270 High-voltage-test-technique-partial discharge measurement

    “IEC 60270 High-voltage-test-technique-partial discharge measurement”

    work page

  21. [21]

    IEC-TS-60034-27-5-2021, Rotating electrical machine- Part 27-5: Offline measurement of partial discharge inception voltage on winding insulation under repetitive impulse voltage

    “IEC-TS-60034-27-5-2021, Rotating electrical machine- Part 27-5: Offline measurement of partial discharge inception voltage on winding insulation under repetitive impulse voltage.”

    work page 2021

  22. [22]

    Thalamo-cortical evoked potentials during stimulation of the dentato- rubro-thalamic tract demonstrate synaptic filtering,

    C. R. Conner, K. J. Forseth, A. M. Lozano, R. Ritter, and A. J. Fenoy, “Thalamo-cortical evoked potentials during stimulation of the dentato- rubro-thalamic tract demonstrate synaptic filtering,” Neurotherapeutics, vol. 21, no. 1, p. e00295, Jan. 2024, doi: 10.1016/j.neurot.2023.10.005

    work page doi:10.1016/j.neurot.2023.10.005 2024

  23. [23]

    Frequency - Based Density Estimation and Identification of Partial Discharges Signal in High-Voltage Generators via Gaussian Mixture Models,

    K. Romphuchaiyapruek and S. Wattanawongpitak, “Frequency - Based Density Estimation and Identification of Partial Discharges Signal in High-Voltage Generators via Gaussian Mixture Models,” Eng, vol. 6, no. 4, p. 64, Mar. 2025, doi: 10.3390/eng6040064

    work page doi:10.3390/eng6040064 2025

  24. [24]

    C. R. L. P. N. Jeukens et al., “A New Algorithm for Automatically Calculating Noise, Spatial Resolution, and Contrast Image Quality Metrics: Proof-of-Concept and Agreement With Subjective Scores in Phantom and Clinical Abdominal CT,” Invest. Radiol., vol. 58, no. 9, pp. 649–655, Sep. 2023, doi: 10.1097/RLI.0000000000000954

    work page doi:10.1097/rli.0000000000000954 2023

  25. [25]

    Partial discharges at a triple junction metal/solid insulator/gas and simulation of inception voltage,

    C. Tran Duy et al., “Partial discharges at a triple junction metal/solid insulator/gas and simulation of inception voltage,” J. Electrost., vol. 66, no. 5– 6, pp. 319–327, May 2008, doi: 10.1016/j.elstat.2008.01.011

    work page doi:10.1016/j.elstat.2008.01.011 2008

  26. [26]

    A Comparative Study of Time -Evolution Characteristics of Single and Double Cavity Partial Discharges,

    I. K. Kyere, C. Nyamupangedengu, and A. G. Swanson, “A Comparative Study of Time -Evolution Characteristics of Single and Double Cavity Partial Discharges,” Energies, vol. 17, no. 8, p. 1905, Apr. 2024, doi: 10.3390/en17081905

    work page doi:10.3390/en17081905 1905

  27. [27]

    Dataset of phase -resolved images of internal, corona, and surface partial discharges in electrical generators,

    J. D. Z. Henao, A. Segura, A. Tenorio, H. J. Diaz, and A. Paz, “Dataset of phase -resolved images of internal, corona, and surface partial discharges in electrical generators,” Data Brief, vol. 52, p. 109992, Feb. 2024, doi: 10.1016/j.dib.2023.109992

    work page doi:10.1016/j.dib.2023.109992 2024

  28. [28]

    Comparative analysis of image classification algorithms based on traditional machine learning and deep learning,

    P. Wang, E. Fan, and P. Wang, “Comparative analysis of image classification algorithms based on traditional machine learning and deep learning,” Pattern Recognit. Lett. , vol. 141, pp. 61 –67, Jan. 2021, doi: 10.1016/j.patrec.2020.07.042

    work page doi:10.1016/j.patrec.2020.07.042 2021

  29. [29]

    Development of Hypergraph Based Improved Random Forest Algorithm for Partial Discharge Pattern Classification,

    S. Govindarajan, J. A. Ardila -Rey, K. Krithivasan, J. Subbaiah, N. Sannidhi, and M. Balasubramanian, “Development of Hypergraph Based Improved Random Forest Algorithm for Partial Discharge Pattern Classification,” IEEE Access , vol. 9, pp. 96 –109, 2021, doi: 10.1109/ACCESS.2020.3047125

    work page doi:10.1109/access.2020.3047125 2021

  30. [30]

    Random Forest Based Optimal Feature Selection for Partial Discharge Pattern Recognition in HV Cables,

    X. Peng et al., “Random Forest Based Optimal Feature Selection for Partial Discharge Pattern Recognition in HV Cables,” IEEE Trans. Power Deliv., vol. 34, no. 4, pp. 1715 –1724, Aug. 2019, doi: 10.1109/TPWRD.2019.2918316

    work page doi:10.1109/tpwrd.2019.2918316 2019

  31. [31]

    Deep learning,

    Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, May 2015, doi: 10.1038/nature14539

    work page doi:10.1038/nature14539 2015

  32. [32]

    Rethinking the Inception Architecture for Computer Vision,

    C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the Inception Architecture for Computer Vision,” presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 2818 –2826. Accessed: Feb. 2 1, 2025. [Online]. Available: https://www.cv- foundation.org/openaccess/content_cvpr_2016/html/Szege...

    work page 2016

  33. [33]

    Going Deeper With Convolutions,

    C. Szegedy et al., “Going Deeper With Convolutions,” presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 1 –9. Accessed: Feb. 21, 2025. [Online]. Available: https://www.cv- foundation.org/openaccess/content_cvpr_2015/html/Szegedy_Going_Deeper_ With_2015_CVPR_paper.html

    work page 2015

  34. [34]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA: IEEE, Jun. 2016, pp. 770–778. doi: 10.1109/CVPR.2016.90

    work page doi:10.1109/cvpr.2016.90 2016