Pith. sign in

REVIEW 4 major objections 6 minor 57 references

Pattern-Based Phase-Separation of Tracer and Dispersed Phase Particles in Two-Phase Defocusing Particle Tracking Velocimetry

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A convolutional neural network can separate tracer particles from bubbles or droplets in defocusing particle tracking velocimetry by reading defocused-image patterns, reaching 95-100% classification accuracy with a GAN-based auto-labeling…

desk verdict Useful integration of CNN detection and GAN auto-labeling for phase separation in two-phase DPTV, but the headline accuracy claim outruns the evidence because real two-phase classification is never measured. read the letter →

arxiv 2506.18157 v1 pith:D5LWMULS submitted 2025-06-22 cs.CV physics.app-phphysics.flu-dyn

classification cs.CVphysics.app-phphysics.flu-dyn
keywords defocusingparticletrackingvelocimetrytwo-phaseflowphaseseparationconvolutionalneuralnetworkgenerativeadversarialinterferencefringepatternimageclassificationauto-labeledtrainingdata
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 sets out to show that dispersed two-phase defocusing particle tracking velocimetry (DPTV) can separate tracer particles from bubbles or droplets in post-processing, using nothing but the visual pattern of each defocused particle image. It argues that bubbles and droplets display regular, unidirectional interference fringes from multiple glare points, while tracers show empty or speckled patterns, and that convolutional neural networks can learn this distinction. To overcome the lack of labeled training data, it introduces a generative-adversarial-network-based auto-labeling pipeline that turns unlabeled single-phase recordings into large, experiment-specific training sets. Across six test sets including synthetic images and real single- and two-phase flows, the networks reach 95-100% classification accuracy on familiar and many unfamiliar conditions. A sympathetic reader would care because, if this holds, a single camera and standard illumination can give simultaneous 3D velocity fields of both phases without wavelength filtering or large size differences.

What carries the argument

The load-bearing object is the defocused particle image pattern: a regular, unidirectional interference fringe pattern produced by two or more glare points on a smooth transparent bubble or droplet, versus empty or irregular speckle patterns from opaque or rough tracers. Two machinery pieces carry the argument: (1) CNN object detectors, specifically Faster R-CNN and YOLOv4 (full and tiny), trained to output a bounding box and a phase class for each particle image; (2) a two-GAN auto-labeling pipeline in which one GAN generates tracer images and a second generates dispersed-phase images, each snippet gets a bounding box from the Hough transform, and the snippets are pasted into synthetic frames with known class labels. The networks are trained on roughly half a million generated particle images and evaluated with class-agnostic average precision, truncated average precision, classification accuracy, and false-positive class bias. The physical scattering model, expressed through glare-point counting, is what makes the pattern a reliable label rather than an arbitrary visual artifact.

What would settle it

Generate a synthetic two-phase test set with the paper's forward model in which tracers are modeled as smooth transparent particles with two ordered glare points, so their defocused images contain the same unidirectional fringes as droplets; if the CNN still classifies the two classes with near-95-100% accuracy, the separation is not actually driven by the fringe pattern, whereas chance-level accuracy would confirm that the pattern is the mechanism.

Watch

Extended reading notes

Core claim

The central claim is that phase identity in DPTV is encoded in the defocused particle image itself: transparent smooth particles such as bubbles and droplets scatter light from multiple glare points lying in one scattering plane, producing unidirectional interference fringes, whereas opaque or rough tracers yield either empty or speckled images. The paper shows that object-detection CNNs (Faster R-CNN and two YOLOv4 variants) trained on auto-labeled GAN-generated images can exploit this pattern difference to detect and classify particle images, with classification accuracy of 95-100% on most datasets. The strongest evidence is the synthetic two-phase set, where idealized, noise-free images force the networks to generalize; high accuracy there indicates the networks learned the fringe-versus-no-fringe physical feature rather than incidental image cues. The paper also demonstrates that the GAN auto-labeling framework removes the manual-annotation bottleneck, with the caveat that smooth transparent tracers such as DEHS can themselves produce fringes, so tracer material must be chosen to keep the pattern distinction valid.

Load-bearing premise

The load-bearing premise is that bubbles and droplets reliably show regular, unidirectional interference fringes in defocused images while the chosen tracers show empty or speckled patterns; if the tracer material also produces clean fringes, the visual cue the classifier relies on disappears.

Editorial extensions

If this is right

  • A single camera with one optical access can produce simultaneous 3D-3C Lagrangian tracks for both the continuous and dispersed phases, since phase separation happens entirely in post-processing.
  • The method stays effective when tracer and dispersed particles have overlapping size distributions, and it does not require high seeding densities or large velocity differences between phases.
  • Because the networks learn a fringe-based physical feature, the approach can be combined with interferometric particle imaging to add size measurement of bubbles or droplets in the same data.
  • The GAN auto-labeling framework can generate experiment-specific training data from roughly 100-1000 single-phase images per class, removing the manual labeling bottleneck for CNN-based DPTV.
  • In domain-shifted test sets, precision remains high even where recall is limited, which suppresses ghost particles in tracking; improving recall is the stated next step for complete void-fraction estimation.

Reading between the lines

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

  • As an editorial extension, the identical two-GAN auto-labeling pipeline could be transferred to astigmatism PTV, where defocused particle images also encode phase-specific appearance and single-camera volumetric tracking is equally valuable.
  • Inferred from the reported false-positive class bias: in applications that count particles rather than track them, such as void-fraction estimation, the dispersed phase would be over-represented unless confidence thresholds are tuned or the training set is rebalanced toward tracers.
  • Inferred from the scattering argument: the same pattern-based separation should work for solid dispersed particles whose surface roughness produces a speckle pattern distinct from tracer fringes, extending the method beyond gas-liquid flows even though the paper does not test this.
  • A practical consequence of the paper's own caveat is that tracer material choice becomes part of experimental design: any new tracer should be screened in single-phase recordings for the presence of regular unidirectional fringes before the two-phase runs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes a post-processing method for phase separation in defocusing particle tracking velocimetry (DPTV). The authors train convolutional object detectors (Faster R-CNN, YOLOv4, and tiny YOLOv4) to distinguish tracer particle images from bubble/droplet particle images on the basis of visual pattern: tracers are assumed to produce empty or speckled defocused particle images, whereas the dispersed phase produces regular, unidirectional interference fringes. To avoid manual labeling, the authors introduce a GAN-based auto-labeling framework in which two class-specific GANs generate synthetic particle images from single-phase calibration recordings, and the generated images are composited with known bounding-box and class labels into training images. The trained detectors are evaluated on six test sets: four real single-phase sets (T1–T4), one synthetic mixed-phase set (T5), and one real mixed-phase set (T6). High detection precision and near-perfect classification accuracy are reported on most familiar and some domain-shifted sets, with notable exceptions on T4 and T5. For the real two-phase set T6, only detection performance is reported because manual class labeling was deemed unreliable.

Significance. If the central claim were fully supported, the work would be a useful practical contribution: it offers a single-camera, post-processing route to phase-specific 3D particle tracking in dispersed two-phase flows without wavelength filtering, and it explicitly addresses the training-data bottleneck via GAN-generated auto-labeled data. The synthetic T5 test is a good positive control: because the synthetic images are nearly noise-free and differ only in the presence of fringes, high classification accuracy for two of the three networks provides credible evidence that at least those networks rely on the intended fringe feature. The GAN auto-labeling workflow is a clear practical step forward for CNN-based DPTV. However, the significance is reduced by the absence of a real mixed-phase classification measurement: the only real two-phase dataset (T6) contributes detection results only, so the paper's headline claim of robust phase separation in real two-phase DPTV is not directly demonstrated. The abstract's '95-100%' accuracy claim is also contradicted by the tables for the domain-shifted sets.

major comments (4)
  1. [§4.1 and Table 2] The central claim that CNN-based phase separation works in real two-phase DPTV is not directly tested. Section 4.1 states that for Exp. T6, the only real two-phase flow dataset, manual class labeling proved unreliable, and Table 2 consequently reports no classification accuracy for T6. The classification accuracies that are reported come from real single-phase sets (T1–T4) and a synthetic mixed-phase set (T5). Single-phase sets cannot exercise genuine class confusion in a mixed-population image, and T5 is an idealized, noise-free case. The abstract's wording that the evaluation 'comprising synthetic two-phase and real single- and two-phase flows, demonstrates high detection precision and classification accuracy' is therefore overstated. To support the central claim, the authors should either provide a real mixed-phase classification result (for example, from carefully reviewed labels, multiple annotators, or a composited image set built from real single-phase images) or explicitly restrict the stated scope of the claim to the data actually tested.
  2. [Abstract, §4.3, Tables 1 and 2] The abstract's '95-100%' classification accuracy and 'even under domain shifts' claim is not supported by the reported numbers. In Table 2, Faster R-CNN achieves 0.754 classification accuracy on T4, YOLOv4 achieves 0.814 on T4 and 0.824 on T5, and only tiny YOLOv4 remains above 0.95 on all real sets. Detection performance also degrades under domain shift: Table 1 shows average precisions of 0.667–0.729 on T2 and 0.765–0.850 on T4. The authors should revise the summary claims to report these exceptions and discuss why the two-stage detector and the full YOLOv4 behave differently, especially the YOLOv4 drop on the synthetic T5 set relative to its strong real-domain performance.
  3. [§4.1 and §4.2] The assertion that the detector 'processes each instance independently, without relying on the overall image composition' is not established by the evaluation. Faster R-CNN and YOLOv4 operate on full images with global receptive fields, and the single-phase test sets T1–T4 provide no mixed-class context that could reveal context-dependent behavior. The paper's argument would be strengthened by a control experiment in which real single-phase particle images are composited into balanced mixed-phase images (with ground-truth classes known from the source sets), or by per-image class-conditional analysis and/or saliency maps showing that classification decisions are driven by the local particle-image pattern rather than background intensity, particle density, or other global cues.
  4. [Table 2 and §4.1] Classification accuracy is computed only over true-positive detections and is reported at a single operating point (the maximum F1-score) without confidence intervals or the per-cell number of true-positive detections. Given that some test sets are small (e.g., T3 contains 202 labeled particle images), the reported accuracies may have wide uncertainty, particularly for the T3 and T6 rows. Reporting the underlying confusion matrices and detection counts would make the differences between networks (e.g., Faster R-CNN's 0.754 vs tiny YOLOv4's 0.966 on T4) more interpretable.
minor comments (6)
  1. [§1 and §2] The verb 'utilities' appears twice ('This work utilities object detection networks' in §1 and 'utilities the pattern' in §2); it should be 'utilizes'.
  2. [Figure 6 caption] The caption reads 'T67' in '(f)'; this should be 'Exp. T6'.
  3. [Figure 7 caption] Both subfigure captions read 'GAN Tracers'; the second subfigure should be labeled 'GAN Dispersed Phase' to match the text describing the two GANs.
  4. [Table 5 header] The table header 'IoU≥0.5Faster R-CNN YOLOv4 YOLOv4' is difficult to parse; the columns for 'Tiny YOLOv4' and 'YOLOv4' should be clearly separated and labeled.
  5. [§5] The sentence 'while the CNNs demonstrated very high TAP with, this performance was achieved only within limited maximum recall ranges' appears to contain a missing word after 'with'; the sentence should be completed.
  6. [Eq. (4)] The normalized truncated average precision divides by max(R) but the text does not specify how max(R) is obtained or how the case max(R)=0 should be handled; this is a minor formal point, but a sentence of clarification would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the evaluation uses held-out real and synthetic test sets, and no predicted quantity reduces to a fitted input or to a self-citation chain.

full rationale

The paper's derivation chain is not circular. The phase-distinction target is an externally stated physical hypothesis, and the detector is trained on GAN-generated images with labels assigned by source GAN, not by the detector itself. All classification evaluations use held-out test sets (T1–T4) that were never used for GAN or detector training; T2 and T4 come from independent experimental groups (wet clutch and spray), providing external benchmark support. T5 is a synthetic forward-model test and is explicitly framed as a controlled probe of the learned fringe feature, not as real-flow validation. The paper itself flags the key evidence gap: in Section 4.1 it states that for the only real two-phase set T6, "manual class labeling of PIs proved unreliable" and therefore "only the detection but not the classification performance was assessed for this dataset." This weakens the abstract's wording but is a missing measurement, not a circular reduction. Similarly, Section 2 acknowledges that smooth transparent tracers such as DEHS can also produce fringes, limiting the pattern assumption; again a scope limitation rather than a self-referential step. The assertion that the detector "processes each instance independently, without relying on the overall image composition" is an untested methodological assumption, but it is not an input-output equivalence. No equation in the paper defines a predicted quantity in terms of the fitted inputs, and no load-bearing argument reduces to a self-citation; the self-citations are to data sources, a forward-model tool, and prior detector settings, none of which force the reported accuracy. The paper is therefore self-contained in its evaluation, and the reported results stand as empirical measurements rather than consequences of an ansatz or a fit. The lack of a classification accuracy for the real two-phase dataset is a genuine limitation for the strongest real-flow claim, but it is a completeness issue, not circularity.

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

The paper introduces no new physical entities or fitted scientific parameters. The neural network weights are fitted to data, but that is model training, not a free parameter in a derivation of a physical result. The core assumptions are physical distinctness of fringe patterns and the transferability of GAN-generated training data.

assumptions (3)
  • domain assumption Bubbles and droplets consistently produce regular, unidirectional interference fringes in defocused images, while tracer particles (with appropriate material choice) do not.
    Section 2 posits this physical distinction based on glare-point scattering theory (van de Hulst), which underlies the entire classification approach. The paper itself notes that transparent tracers like DEHS can also produce fringes.
  • domain assumption GAN-generated particle images are sufficiently realistic that a detector trained on them transfers to real experimental images.
    The auto-labeling framework (Section 3.1) relies on this transfer. No direct comparison is made between training on GAN data versus training on real extracted particle images or physics-based synthetic images.
  • domain assumption Manual labeling of single-phase test images yields reliable ground truth for classification evaluation.
    Section 4.1 assumes class labels are 'known with absolute certainty' in single-phase images. For the real two-phase set T6, manual labels were considered unreliable and classification was not scored.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pattern-Based Phase-Separation of Tracer and Dispersed Phase Particles in Two-Phase Defocusing Particle Tracking Velocimetry." pith.science (2026). https://pith.science/paper/D5LWMULS

@misc{pith2026250618157,
  author       = {Pith},
  title        = {Pith review of: Pattern-Based Phase-Separation of Tracer and Dispersed Phase Particles in Two-Phase Defocusing Particle Tracking Velocimetry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D5LWMULS}},
  note         = {Machine review of arXiv:2506.18157}
}
read the original abstract

This work investigates the feasibility of a post-processing-based approach for phase separation in defocusing particle tracking velocimetry for dispersed two-phase flows. The method enables the simultaneous 3D localization determination of both tracer particles and particles of the dispersed phase, using a single-camera setup. The distinction between phases is based on pattern differences in defocused particle images, which arise from distinct light scattering behaviors of tracer particles and bubbles or droplets. Convolutional neural networks, including Faster R-CNN and YOLOv4 variants, are trained to detect and classify particle images based on these pattern features. To generate large, labeled training datasets, a generative adversarial network based framework is introduced, allowing the generation of auto-labeled data that more closely reflects experiment-specific visual appearance. Evaluation across six datasets, comprising synthetic two-phase and real single- and two-phase flows, demonstrates high detection precision and classification accuracy (95-100%), even under domain shifts. The results confirm the viability of using CNNs for robust phase separation in disperse two-phase DPTV, particularly in scenarios where traditional wavelength-, size-, or ensemble correlation-based methods are impractical.

Figures

Figures reproduced from arXiv: 2506.18157 by the authors.

Figure 1
Figure 1. Three different types of PIs: PIs with an empty pattern (left), PIs with a regular unidirectional pattern (middle) and PIs with a irregular speckle pattern (right). The particle type, the scattering characteristic and simulated and real examples are shown. The simulated examples where created using the forward model from Sax et al. 46 and the glare point based scattering characteristic. For the rough opaque particle… view at source ↗
Figure 2
Figure 2. Auto-labeling approach and PI detection. Two data acquisition experiments are conducted, each involving a single phase. PIs are extracted from the resulting image sets using conventional algorithms. These extracted PIs are then used to train GANs, to generate mixed-phase image sets. The mixed-phase sets are subsequently used to train an object detection CNN. Finally, the trained CNN is applied to the actual two-phas… view at source ↗
Figure 3
Figure 3. Generation of automatically labeled training data for object detection networks. Two GANs - one for tracer particles and one for the dispersed phase - independently generate small image snippets, each containing a single PI. For each snippet, a BB is placed over the PI. The image snippets and corresponding BBs are then randomly resized and inserted into an empty image. To train the GANs, single-phase data acquisitio… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: PIs sampled from the Exp. G1 7 (a), Exp. G2 52 (b), Exp. G3 53 (c) and Exp. G4 7 (d). These PIs were used to train the GANs for tracer particles (a)&(b) and for particles of the dispersed phase (c)&(d). (a) GAN Tracers (b) GAN Dispersed Phase [PITH_FULL_IMAGE:figures/…
Figure 5
Figure 5. Figure 5: GAN-generated images of PIs of tracer particles (a) and particles of the dispersed phase (b). All images have dimensions of 64×64 px. the dispersed phase, which was trained on the smaller of the two datasets, exhibited a divergence between the generator and discriminat…
Figure 6
Figure 6. Figure 6: Exemplary images from the six test datasets: T17 (a) and T262 (b) show tracer particles in a single phase flow. T353 (c) and T463 (d) show only bubbles and droplets but no tracers. T5 (e) shows a synthetic image of a two phase flow of bubbles and tracers and T67 (f) sh…
Figure 7
Figure 7. Figure 7: Generator score (G-score) and discriminator score (D-score) during the training of the two GANs. Two GANs were trained: One for the generation of PIs representing tracers (a) and another to represent particles of the dispersed phase (b). In both cases the D-Score conve…
Figure 8
Figure 8. Figure 8: Total training loss of Faster R-CNN (a) and YOLOv4 and tiny YOLOv4 (b). False Positive Class Bias For the three object detection networks, Faster R-CNN, YOLOv4, and Tiny YOLOv4, the false positive class bias was computed and is presented in [PITH_FULL_IMAGE:figures/fu…
Figure 9
Figure 9. Figure 9: False positive class biases for Exps. T1, T2, T3, T4, T5, T6 over the applied confidence score threshold. The black dashed line indicates zero bias. The bias becomes +1 if all FP are predicted to be tracers and -1 if predicted as dispersed phase. 19/19 [PITH_FULL_IMAG…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 46 canonical work pages

  1. [3]

    & Henry, C.Understanding Turbulent Systems - Progress in Particle Dynamics Modeling (Springer Nature Switzerland AG, 2025)

    Minier, J.-P., Ferrand, M. & Henry, C.Understanding Turbulent Systems - Progress in Particle Dynamics Modeling (Springer Nature Switzerland AG, 2025)

  2. [4]

    Computational flow modelling and design of bubble column reactors.Chem

    Joshi, J. Computational flow modelling and design of bubble column reactors.Chem. Eng. Sci.56, 5893–5933, DOI: https://doi.org/10.1016/S0009-2509(01)00273-1 (2001). Proceedings of the 5th International Conference on Gas-Liquid and Gas-Liquid-Solid Reactor Engineering. 13/19

  3. [5]

    & Hofmann, H

    Grienberger, J. & Hofmann, H. Investigations and modelling of bubble columns.Chem. Eng. Sci.47, 2215–2220, DOI: https://doi.org/10.1016/0009-2509(92)87037-Q (1992). Twelfth International Symposium on Chemical Reaction Engineering Today

  4. [6]

    & Kriegseis, J

    Leister, R., Brümmer, A. & Kriegseis, J. Laser-optical shear-flow analysis across the annular gap of a simplified displacement compressor model.IOP Conf. Series: Mater. Sci. Eng.1267, 012003, DOI: 10.1088/1757-899X/1267/1/ 012003 (2022)

  5. [7]

    & Kriegseis, J

    Lange, H., Sax, C., Brümmer, A. & Kriegseis, J. Astigmatism quantification for depth localization of bubbles and tracers across curved surfaces.IOP Conf. Series: Mater. Sci. Eng.1322, DOI: 10.1088/1757-899X/1322/1/012005 (2024)

  6. [8]

    D., Kourmatzis, A

    Gounder, J. D., Kourmatzis, A. & Rachid Masri, A. Turbulent piloted dilute spray flames: Flow fields and droplet dynamics. Combust. Flame159, 3372–3397, DOI: https://doi.org/10.1016/j.combustflame.2012.07.014 (2012)

  7. [9]

    & Taylor, A

    Sahu, S., Hardalupas, Y . & Taylor, A. M. K. P. Interaction of droplet dispersion and evaporation in a polydispersed spray.J. Fluid Mech.846, 37–81, DOI: 10.1017/jfm.2018.247 (2018)

  8. [10]

    & Maliska, C

    Cerqueira, R., Paladino, E., Ynumaru, B. & Maliska, C. Image processing techniques for the measurement of two- phase bubbly pipe flows using particle image and tracking velocimetry (piv/ptv).Chem. Eng. Sci.189, 1–23, DOI: https://doi.org/10.1016/j.ces.2018.05.029 (2018)

Show all 57 references
  1. [11]

    & Kouda, T

    Kitagawa, A., Hagiwara, Y . & Kouda, T. Ptv investigation of phase interaction in dispersed liquid-liquid two-phase turbulent swirling flow.Exp. Fluids42, 871–880, DOI: 10.1007/s00348-007-0291-5 (2007)

  2. [12]

    Kiger, K. T. & Pan, C. Piv technique for the simultaneous measurement of dilute two-phase flows.J. Fluids Eng.122, 811–818, DOI: 10.1115/1.1314864 (2000). https://doi.org/10.1115/1.1314864

  3. [13]

    & Fan, L.-S

    Chen, R. & Fan, L.-S. Particle image velocimetry for characterizing the flow structure in three-dimensional gas-liquid-solid fluidized beds.Chem. Eng. Sci.47, 3615–3622, DOI: 10.1016/0009-2509(92)85077-O (1992)

  4. [14]

    L., Easson, W

    Jakobsen, M. L., Easson, W. J., Greated, C. A. & Glass, D. H. Particle image velocimetry: simultaneous two-phase flow measurements.Meas. Sci. Technol.7, 1270, DOI: 10.1088/0957-0233/7/9/012 (1996)

  5. [15]

    & Lecordier, B

    Kosiwczuk, W., Cessou, A., Trinité, M. & Lecordier, B. Simultaneous velocity field measurements in two-phase flows for turbulent mixing of sprays by means of two-phase piv.Exp. Fluids39, 895–908, DOI: 10.1007/s00348-005-0027-3 (2005)

  6. [16]

    & Longmire, E

    Khalitov, D. & Longmire, E. Simultaneous two-phase piv by two-parameter phase discrimination.Exp. Fluids32, 252–268, DOI: 10.1007/s003480100356 (2002)

  7. [17]

    & Westerweel, J

    Delnoij, E., Kuipers, J., van Swaaij, W. & Westerweel, J. Measurement of gas–liquid two-phase flow in bubble columns using ensemble correlation piv.Chem. Eng. Sci.55, 3385–3395, DOI: 10.1016/S0009-2509(99)00595-3 (2000)

  8. [18]

    & Socolofsky, S

    Seol, D. & Socolofsky, S. Vector post-processing algorithm for phase discrimination of two-phase piv.Exp. Fluids45, 223–239, DOI: 10.1007/s00348-008-0473-9 (2008)

  9. [19]

    & Hirata, M

    Nishino, K., Kasagi, N. & Hirata, M. Three-dimensional particle tracking velocimetry based on automated digital image processing.J. Fluids Eng.111, 384–391, DOI: 10.1115/1.3243657 (1989). https://asmedigitalcollection.asme.org/ fluidsengineering/article-pdf/111/4/384/5900156/384_1.pdf

  10. [20]

    & Papantoniou, D

    Maas, H., Gruen, A. & Papantoniou, D. Particle tracking velocimetry in three-dimensional flows.Exp. Fluids15, 133–146, DOI: https://doi.org/10.1007/BF00190953 (1993)

  11. [21]

    & Wieneke, B

    Schanz, D., Schröder, A., Gesemann, S., Michaelis, D. & Wieneke, B. ‘shake the box’: A highly efficient and accurate tomographic particle tracking velocimetry (tomo-ptv) method using prediction of particle positions. InPIV13; 10th International Symposium on Particle Image V el...

  12. [22]

    & Gesemann, S

    Schanz, D. & Gesemann, S. . S. Shake-the-box: Lagrangian particle tracking at high particle image densities.Exp Fluids 57, 70DOI: 10.1007/s00348-016-2157-1 (2016). 23.Willert, C. E. & Gharib, M. Three-dimensional particle imaging with a single camera.Exp. Fluids12, 353–358 (1992)

  13. [24]

    & Adrian, R

    Olsen, M. & Adrian, R. Out-of-focus effects on particle image visibility and correlation in microscopic particle image velocimetry.Exp. Fluids 29DOI: 10.1007/s003480070018 (2000)

  14. [25]

    & Kähler, C

    Cierpka, C., Segura, R., Hain, R. & Kähler, C. J. A simple single camera 3c3d velocity measurement technique without errors due to depth of correlation and spatial averaging for microfluidics.Meas. Sci. Technol.21, 045401, DOI: 10.1088/0957-0233/21/4/045401 (2010)

  15. [26]

    & Verkman, A

    Kao, H. & Verkman, A. Tracking of single fluorescent particles in three dimensions: use of cylindrical optics to encode particle position.Biophys. J.67, 1291–1300, DOI: https://doi.org/10.1016/S0006-3495(94)80601-0 (1994). 14/19

  16. [27]

    Neural Inf

    Lecun, Y .et al.Handwritten digit recognition with a backpropagation network.Adv. Neural Inf. Process. Syst. volume 2 1990(1990)

  17. [28]

    & Haffner, P

    Lecun, Y ., Bottou, L., Bengio, Y . & Haffner, P. Gradient-based learning applied to document recognition.Proc. IEEE86, 2278–2324 (1998)

  18. [29]

    & Mäder, P

    Cierpka, C., König, J., Chen, M., Boho, D. & Mäder, P. On the use of machine learning algorithms for the calibration of astigmatism ptv.Proc. 13th Int. Symp. on Part. Image V elocim.(2019)

  19. [30]

    & Krevor, S

    Franchini, S. & Krevor, S. Cut, overlap and locate: a deep learning approach for the 3d localization of particles in astigmatic optical setups.Exp. Fluids61(2020)

  20. [31]

    König, J.et al.On the use of a cascaded convolutional neural network for three-dimensional flow measurements using astigmatic ptv.Meas. Sci. Technol.31, 074015 (2020)

  21. [32]

    Barnkob, R.et al.Defocus particle tracking: a comparison of methods based on model functions, cross-correlation, and neural networks.Meas. Sci. Technol.32, 094011 (2021)

  22. [33]

    Dreisbach, M.et al.Particle detection by means of neural networks and synthetic training data refinement in defocusing particle tracking velocimetry.Meas. Sci. Technol.33, 124001 (2022)

  23. [34]

    & Kriegseis, J

    Sax, C., Dreisbach, M., Leister, R. & Kriegseis, J. Deep learning and hybrid approach for particle detection in defocusing particle tracking velocimetry.Meas. Sci. Technol.34, 095909, DOI: 10.1088/1361-6501/acd4b4 (2023)

  24. [35]

    Zhang, X.et al.3d positioning and autofocus of the particle field based on the depth-from-defocus method and the deep networks.Mach. Learn. Sci. Technol.4, 025030, DOI: 10.1088/2632-2153/acdb2e (2023)

  25. [36]

    & Cierpka, C

    Ratz, M., Sachs, S., König, J. & Cierpka, C. A deep neural network architecture for reliable 3d position and size determination for lagrangian particle tracking using a single camera.Meas. Sci. Technol.34, 105203, DOI: 10.1088/ 1361-6501/ace070 (2023)

  26. [37]

    & Sun, J

    Ren, S., He, K., Girshick, R. & Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis machine intelligence39, 1137–1149 (2017)

  27. [38]

    & Farhadi, A

    Redmon, J., Divvala, S., Girshick, R. & Farhadi, A. You only look once: Unified, real-time object detection. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 779–788, DOI: 10.1109/CVPR.2016.91 (2016)

  28. [39]

    & Farhadi, A

    Redmon, J. & Farhadi, A. Yolo9000: Better, faster, stronger. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 6517–6525, DOI: 10.1109/CVPR.2017.690 (2017). 40.Redmon, J. & Farhadi, A. Yolov3: An incremental improvement (2018). 1804.02767. 41.Goodfellow...

  29. [42]

    & Liao, H.-y

    Bochkovskiy, A., Wang, C.-Y . & Liao, H.-y. Yolov4: Optimal speed and accuracy of object detection, DOI: 10.48550/ arXiv.2004.10934 (2020)

  30. [43]

    & Frohn, A

    König, G., Anders, K. & Frohn, A. A new light-scattering technique to measure the diameter of periodically generated moving droplets.J. Aerosol Sci.17, 157–167, DOI: https://doi.org/10.1016/0021-8502(86)90063-7 (1986)

  31. [44]

    R., Skippon, S

    Glover, A. R., Skippon, S. M. & Boyle, R. D. Interferometric laser imaging for droplet sizing: a method for droplet-size measurement in sparse spray systems.Appl. Opt.34, 8409–8421 (1995). https://doi.org/10.1364/AO.34.008409

  32. [45]

    & Maeda, M

    Niwa, Y ., Kamiya, Y ., Kawaguchi, T. & Maeda, M. Bubble sizing by interferometric laser imaging.10th Int. Symp. on Appl. Laser Tech. to Fluid Mech.(2000)

  33. [46]

    & Kriegseis, J

    Sax, C., Griesmaier, R. & Kriegseis, J. Particle localization with dptv and sizing with ipi using a forward model and optimization.Meas. Sci. Technol.Manuscript under Review(2025)

  34. [47]

    Laser Speckle and Related Phenomena(Springer Nature, Topics in Applied Physics, 1975)

    Dainty, J.et al. Laser Speckle and Related Phenomena(Springer Nature, Topics in Applied Physics, 1975). https: //doi.org/10.1007/978-3-662-43205-1

  35. [48]

    van de Hulst, H. C. & Wang, R. T. Glare points.Appl. optics30, 4755–4763, DOI: https://doi.org/10.1364/AO.30.004755 (1991). 49.Atherton, T. J. & Kerbyson, D. J. The coherent circle hough transform.Br. Mach. Vis. Conf.27.1–27.10 (1993)

  36. [50]

    & Kerbyson, D

    Atherton, T. & Kerbyson, D. Using phase to represent radius in the coherent circle hough transform. InIEE Colloquium on Hough Transforms, 5/1–5/4 (1993). 51.Atherton, T. & Kerbyson, D. Size invariant circle detection.Image Vis. Comput.17, 795–803 (1999). 15/19

  37. [52]

    & Kriegseis, J

    Pasch, S., Lange, H., Leister, R. & Kriegseis, J. Flow measurements above a dbd plasma actuator array by means of defocusing ptv. InProceedings of the International Symposium on the Application of Laser and Imaging Techniques to Fluid Mechanics, vol. 21, 1–19, DOI: 10.55037/lx...

  38. [53]

    & Kriegseis, J

    Sax, C., Dreisbach, M. & Kriegseis, J. Interferometric particle imaging for particle sizing in the front-, side-, and back-scatter region (2023). https://arxiv.org/abs/2303.16013, 2303.16013

  39. [54]

    Matlab - gan

    MathWorks, I. Matlab - gan. https://www.mathworks.com/help/deeplearning/ug/train-generative-adversarial-network.html (2020). [Online; last accessed 25-May-2025]

  40. [55]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770–778, DOI: 10.1109/CVPR.2016.90 (2016)

  41. [56]

    Software available from https://www.tensorflow.org/

    Abadi, M.et al.TensorFlow: Large-scale machine learning on heterogeneous systems (2015). Software available from https://www.tensorflow.org/

  42. [57]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Spatial pyramid pooling in deep convolutional networks for visual recognition.IEEE Transactions on Pattern Analysis Mach. Intell.37, 1904–1916, DOI: 10.1109/TPAMI.2015.2389824 (2015)

  43. [58]

    & Jia, J

    Liu, S., Qi, L., Qin, H., Shi, J. & Jia, J. Path aggregation network for instance segmentation. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8759–8768, DOI: 10.1109/CVPR.2018.00913 (2018)

  44. [59]

    https://doi.org/10.48550/ arXiv.1911.11929, 1911.11929

    Wang, C.-Y .et al.Cspnet: A new backbone that can enhance learning capability of cnn (2019). https://doi.org/10.48550/ arXiv.1911.11929, 1911.11929

  45. [60]

    Lin, T.-Y .et al.Microsoft coco: Common objects in context // microsoft coco: Common objects in context.Eur. Conf. on Comput. Vision, 2014(2015). http://dx.doi.org/10.48550/arXiv.1405.0312

  46. [61]

    & Belongie, S

    Cui, Y ., Song, Y ., Sun, C., Howard, A. & Belongie, S. Large scale fine-grained categorization and domain-specific transfer learning.IEEE/CVF Conf. on Comput. Vis. Pattern Recognit.4109–4118 (2018). http://dx.doi.org/10.1109/CVPR.2018. 00432

  47. [62]

    & Kriegseis, J

    Leister, R., Fuchs, T., Mattern, P. & Kriegseis, J. Flow-structure identification in a radially grooved open wet clutch by means of defocusing particle tracking velocimetry.Exp. Fluids62(2021)

  48. [63]

    Master’s thesis, Karlsruhe Institute of Technology (2023)

    Pöppe, D.Evaluating the Applicability of Flow Measurement Techniques for Characterizing the Microdroplet Flow of the Aerosol-on-Demand Printer. Master’s thesis, Karlsruhe Institute of Technology (2023)

  49. [64]

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization.CoRRabs/1412.6980(2014). https://doi.org/10. 48550/arXiv.1412.6980. 16/19 6 Appendix GAN Training Configuration and Training Scores This section provides further details on the GAN architecture (Tab. 3) and the ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.