pith. sign in

arxiv: 2510.06194 · v2 · pith:D5UHF3HFnew · submitted 2025-10-07 · ✦ hep-ex · astro-ph.IM· cs.CV

Overlap-aware segmentation for topological reconstruction of obscured objects

J. Schueler , H. M. Ara\'ujo , S. N. Balashov , J. E. Borg , C. Brew , F. M. Brunbauer , C. Cazzaniga , A. Cottle
show 25 more authors
D. Edgeman C. D. Frost F. Garcia D. Hunt M. Kastriotou P. Knights H. Kraus A. Lindote M. Lisowska D. Loomba E. Lopez Asamar P. A. Majewski T. Marley C. McCabe L. Millins R. Nandakumar T. Neep F. Neves K. Nikolopoulos E. Oliveri A. Roy T. J. Sumner E. Tilly W. Thompson M. A. Vogiatzi
This is my paper

Pith reviewed 2026-05-21 21:20 UTC · model grok-4.3

classification ✦ hep-ex astro-ph.IMcs.CV
keywords overlap-aware segmentationMIGDAL experimentelectron recoilnuclear recoilMigdal effectsegmentation regressionweighted losstopological reconstruction
0
0 comments X

The pith

A new overlap-weighted loss in segmentation regression cuts reconstruction errors for faint electron tracks from 41% to 13%.

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

The paper introduces OASIS, a segmentation-regression framework whose loss function assigns higher weight to pixels where objects overlap. This targets the specific problem of recovering a faint electron recoil track that is often buried inside a much brighter nuclear recoil track inside the MIGDAL optical time projection chamber. Across eight training runs the overlap weighting alone moves median intensity error on the low-energy electrons from -41.1 % to -13.3 %. The same change also improves topological feature recovery. The result matters because it shows how to extract usable signals from regions where attribution is most ambiguous without treating every pixel equally.

Core claim

OASIS demonstrates that the single most effective training modification for low-energy electron reconstruction is the addition of overlap-targeted weights to the loss function. These weights focus learning on the pixels where the faint track overlaps the brighter nuclear recoil, directly addressing the dominant source of intensity and topology errors. Averaged over eight independent training campaigns, the weighted model reduces median intensity reconstruction error from -41.1 % to -13.3 % while preserving performance outside overlap regions.

What carries the argument

Overlap-targeted loss weighting inside a segmentation-regression network, which raises the penalty on pixels belonging to both the faint electron track and the brighter nuclear recoil so the model learns to apportion intensity correctly in ambiguous zones.

If this is right

  • Pixel intensities of the low-energy electron can be recovered with roughly one-third the previous median error.
  • Topological features such as track length and direction become measurable even when the electron track is visually submerged.
  • The same weighting trick applies to any segmentation-regression task in which one object is orders of magnitude fainter than another and they share pixels.
  • Event selection and background rejection in the MIGDAL detector improve because the reconstructed electron properties are less biased.

Where Pith is reading between the lines

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

  • The same overlap-weighting idea could be tested on other rare-event searches where a faint ionization signal sits on top of a brighter track, such as in directional dark-matter detectors.
  • One could generate synthetic overlap datasets with known ground-truth intensities to isolate whether the weighting introduces any systematic bias in non-overlapping pixels.
  • If the method generalizes, it may reduce the need for hardware upgrades that increase spatial resolution simply to avoid overlaps.

Load-bearing premise

The main source of error when reconstructing the faint electron track is its pixel-by-pixel overlap with the brighter nuclear recoil, and raising the loss weight on those shared pixels will reduce the error without creating new inaccuracies outside the overlap zones.

What would settle it

Measure intensity and topology errors on a held-out set of simulated tracks that have identical brightness but zero spatial overlap; if the overlap-weighted model still shows the same 28-percentage-point improvement, the claim that overlap is the dominant error source does not hold.

Figures

Figures reproduced from arXiv: 2510.06194 by A. Cottle, A. Lindote, A. Roy, C. Brew, C. Cazzaniga, C. D. Frost, C. McCabe, D. Edgeman, D. Hunt, D. Loomba, E. Lopez Asamar, E. Oliveri, E. Tilly, F. Garcia, F. M. Brunbauer, F. Neves, H. Kraus, H. M. Ara\'ujo, J. E. Borg, J. Schueler, K. Nikolopoulos, L. Millins, M. A. Vogiatzi, M. Kastriotou, M. Lisowska, P. A. Majewski, P. Knights, R. Nandakumar, S. N. Balashov, T. J. Sumner, T. Marley, T. Neep, W. Thompson.

Figure 1
Figure 1. Figure 1: Schematic of OASIS. An input intensity-map image of dimension nx × ny containing k distinct object classes (k = 5 in this example) is passed as input into the network. The backbone network, where here we use U–Net, processes the images which are then passed into a segmentation-regression head, shown as the orange block. This maps the backbone network’s output to k intensity maps of dimension nx × ny, each … view at source ↗
Figure 2
Figure 2. Figure 2: Illustration showing the hybrid signal simulation construction of two events: one with a 5.8 keV ER (top row) and the other with a 9.8 keV ER (bottom row). For each event, the hybrid signal (left column) is constructed by stitching a simulated ER track (middle column) with a real, measured NR track (right column). The stitching point is the truth vertex position of the simulated ER and the estimated vertex… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of OASIS’s track reconstruction performance between the unweighted and weighted training campaigns. Left: Median (points) and 25th-75th percentile ranges (error bands) of the percent error of predicted ER track intensity compared to truth, ∆I, versus truth ER energy. Right: Median (points) and 25th-75th percentile ranges (error bands) of pixel IoU between OASIS’s predicted ER and truth versus tr… view at source ↗
Figure 4
Figure 4. Figure 4: Four test set examples comparing OASIS’s ER reconstruction performance to truth. In panels (a)-(c) both the truth and predicted ERs have principal curves shown in white with estimated directional axes shown in black. Panel (d) is an NR-only input event. Predicted ER and truth ER energies for each panel are: (a) EˆER = 5.2 keV, Etruth,ER = 5.2 keV. (b) EˆER = 5.5 keV, Etruth,ER = 5.2 keV. (c) EˆER = 0.89 ke… view at source ↗
Figure 5
Figure 5. Figure 5: Left: Histograms of reconstructed and truth ER intensities for the test set signal sample (blue and orange, respectively), and reconstructed ER intensities for the NR-only sample (green). The black dashed line represents the 3 keV false positive threshold for ER detection in the NR-only sample. Right: Model’s predicted ER track intensity (IˆER) versus truth ER track intensity (Itruth,ER) for all test set E… view at source ↗
Figure 6
Figure 6. Figure 6: Left: Median (points) and inter-quartile range (error bars) of angular consistency ∆ϕ versus |∆I|. Right: Median (points) and inter-quartile range (error bars) or ∆ϕ versus truth ER energy. topologies, ∆ϕ ∈ [0, 90◦ ], as our heuristic for topological similarity. The left panel of [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
read the original abstract

The separation of overlapping objects presents a significant challenge in scientific imaging. While deep learning segmentation-regression algorithms can predict pixel-wise intensities, they typically treat all regions equally rather than prioritizing overlap regions where attribution is most ambiguous. Recent advances in instance segmentation show that weighting regions of pixel overlap in training can improve segmentation boundary predictions in regions of overlap, but this idea has not yet been extended to segmentation regression. We address this with Overlap-Aware Segmentation of ImageS (OASIS): a new segmentation-regression framework with a weighted loss function designed to prioritize regions of object-overlap during training, enabling extraction of pixel intensities and topological features from heavily obscured objects. We demonstrate OASIS in the context of the MIGDAL experiment, which aims to directly image the Migdal effect--a rare process where electron emission is induced by nuclear scattering--in a low-pressure optical time projection chamber. This setting poses an extreme test case, as the target for reconstruction is a faint electron recoil track which is often heavily-buried within the order(s)-of-magnitude brighter nuclear recoil track. Compared to unweighted segmentation regression, we demonstrate OASIS's novel overlap region-targeted loss function weight to be the single most important training weight for improving intensity and topological reconstructions of the low-energy electron tracks that tend to be most dominated by pixel overlap. Averaging over eight training campaigns, we further show the addition of overlap-targeted weights to improve median intensity reconstruction errors from -41.1% to -13.3% for these low-energy electrons. These performance gains demonstrate OASIS as a generalizable methodology for recovering obscured signals in overlap-dominated regions.

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

3 major / 2 minor

Summary. The manuscript introduces OASIS, a segmentation-regression framework that incorporates an overlap-targeted weighted loss function to prioritize ambiguous overlap pixels during training. Applied to the MIGDAL optical TPC data, the method targets reconstruction of faint low-energy electron tracks that are typically obscured by brighter nuclear recoil tracks. The central empirical claim is that adding the overlap weight is the single most important training factor and, when averaged over eight independent training campaigns, reduces median intensity reconstruction error for these electrons from -41.1% to -13.3% relative to the unweighted baseline.

Significance. If the reported improvement is robust, the work would offer a practical, low-overhead modification to existing segmentation-regression pipelines for overlap-dominated scientific imaging. The MIGDAL test case is a stringent one, and a demonstrated gain on held-out experimental data would be of direct interest to the low-energy nuclear recoil community. The absence of variance estimates and detailed loss specifications, however, prevents a firm assessment of whether the result generalizes beyond the specific runs shown.

major comments (3)
  1. Abstract: the central quantitative claim reports median intensity errors improving from -41.1% to -13.3% when averaged over eight training campaigns, yet supplies neither per-campaign values, standard deviations, nor any measure of run-to-run variability. Without these statistics it is impossible to determine whether the observed shift exceeds typical stochastic fluctuations in a regression task on low-statistics tracks.
  2. Abstract (paragraph describing the loss function): the overlap-region loss weight is identified as the single most important hyper-parameter, but the manuscript provides neither its numerical value nor the precise functional form of the weighted loss. Because this weight is the only free parameter highlighted in the method, its omission prevents both reproducibility and any assessment of sensitivity to its choice.
  3. Abstract: the comparison is limited to the unweighted case; no additional baselines (e.g., standard focal loss, boundary-aware losses, or instance-segmentation weighting schemes) are reported. This leaves open whether the observed gain is specific to the proposed overlap weighting or could be obtained by other established re-weighting strategies.
minor comments (2)
  1. The abstract states that the method enables extraction of both pixel intensities and topological features, but the manuscript does not quantify the topological improvement (e.g., track length, curvature, or vertex resolution) with the same numerical detail given for intensity error.
  2. No discussion is provided of possible post-hoc selection of the eight training campaigns or of any data-augmentation or early-stopping choices that might affect the reported medians.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their detailed and constructive review. Their comments have prompted us to enhance the abstract with additional statistical information and clarifications to better support the reproducibility and assessment of our results. Below we respond to each major comment.

read point-by-point responses

  1. Referee: Abstract: the central quantitative claim reports median intensity errors improving from -41.1% to -13.3% when averaged over eight training campaigns, yet supplies neither per-campaign values, standard deviations, nor any measure of run-to-run variability. Without these statistics it is impossible to determine whether the observed shift exceeds typical stochastic fluctuations in a regression task on low-statistics tracks.

    Authors: We agree with the referee that providing measures of variability is important for assessing the robustness of the reported improvement. In the revised manuscript, we have included the standard deviation of the median errors across the eight independent training campaigns in the abstract. Additionally, we have added a table in the supplementary material that lists the per-campaign reconstruction errors for both the baseline and OASIS approaches, allowing readers to evaluate the consistency of the performance gain. revision: yes

  2. Referee: Abstract (paragraph describing the loss function): the overlap-region loss weight is identified as the single most important hyper-parameter, but the manuscript provides neither its numerical value nor the precise functional form of the weighted loss. Because this weight is the only free parameter highlighted in the method, its omission prevents both reproducibility and any assessment of sensitivity to its choice.

    Authors: The precise functional form of the overlap-weighted loss is provided in the Methods section of the manuscript. We concur that including the specific numerical value of the weight in the abstract would improve accessibility and reproducibility. We have therefore revised the abstract to explicitly state the value of the overlap weight and provide a concise description of how it is applied in the loss function. revision: yes

  3. Referee: Abstract: the comparison is limited to the unweighted case; no additional baselines (e.g., standard focal loss, boundary-aware losses, or instance-segmentation weighting schemes) are reported. This leaves open whether the observed gain is specific to the proposed overlap weighting or could be obtained by other established re-weighting strategies.

    Authors: Our study focuses on demonstrating the impact of incorporating overlap-aware weighting into a standard segmentation-regression framework by comparing directly against the unweighted baseline. This choice isolates the contribution of the overlap weighting. While we recognize the value of benchmarking against other established methods such as focal loss or boundary-aware losses, these are primarily developed for classification or different objectives. We have added a paragraph in the discussion section of the revised manuscript to address this point, explaining the rationale for our baseline choice and suggesting that comparisons to alternative re-weighting strategies represent a valuable avenue for future investigation. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical gains on held-out data are independent of inputs

full rationale

The paper presents OASIS as a segmentation-regression framework whose core innovation is an overlap-targeted weighting term in the loss function. The central quantitative claim—an improvement in median intensity reconstruction error from -41.1% to -13.3% when averaged over eight training campaigns—is obtained by training the model on experimental MIGDAL data and evaluating on held-out tracks. No equations are supplied that would make the reported error reduction equivalent to a fitted parameter or to the weighting choice by algebraic construction. No self-citations are used to justify uniqueness or to import an ansatz; the method is described as a direct extension of existing instance-segmentation weighting ideas to regression, with performance verified externally on independent runs rather than derived from prior author results. The derivation chain therefore remains self-contained and falsifiable against the experimental test set.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of an overlap-weighted loss term whose exact functional form and hyper-parameter values are not supplied in the abstract; the method assumes standard deep-learning segmentation-regression can be improved by re-weighting overlap pixels without further theoretical justification.

free parameters (1)
  • overlap-region loss weight
    Hyper-parameter that controls how much extra emphasis is given to overlap pixels; its value is not stated and must be chosen or tuned.
axioms (1)
  • domain assumption Deep-learning segmentation-regression networks can be trained to predict pixel-wise intensities from overlapping objects.
    Invoked when the authors state that standard networks treat all regions equally and that re-weighting overlap improves results.

pith-pipeline@v0.9.0 · 5991 in / 1374 out tokens · 61813 ms · 2026-05-21T21:20:31.263743+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Optical effects in Gaseous Electron Multipliers (GEMs)

    physics.ins-det 2026-04 unverdicted novelty 5.0

    Optical broadening in glass GEMs increases simulated particle track intensity and width by up to 26% and 31%, explaining larger-than-expected signals in GEM-based OTPCs.

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    Melchior et al

    P. Melchior et al. SCARLET: Source separation in multi-band images by Constrained Matrix Factorization.Astronomy and Computing, 24:129, July 2018

    work page 2018

  2. [2]

    Deep learning enables fast and dense single-molecule localization with high accuracy.Nature Methods, 18(9):1082–1090, 2021

    Artur Speiser et al. Deep learning enables fast and dense single-molecule localization with high accuracy.Nature Methods, 18(9):1082–1090, 2021

    work page 2021

  3. [3]

    Boundary-weighted Domain Adaptive Neural Network for Prostate MR Image Segmentation.arXiv e-prints, page arXiv:1902.08128, February 2019

    Qikui Zhu, Bo Du, and Pingkun Yan. Boundary-weighted Domain Adaptive Neural Network for Prostate MR Image Segmentation.arXiv e-prints, page arXiv:1902.08128, February 2019

    work page arXiv 1902

  4. [4]

    DoNet: Deep De-overlapping Network for Cytology Instance Segmentation.arXiv e-prints, page arXiv:2303.14373, March 2023

    Hao Jiang, Rushan Zhang, Yanning Zhou, Yumeng Wang, and Hao Chen. DoNet: Deep De-overlapping Network for Cytology Instance Segmentation.arXiv e-prints, page arXiv:2303.14373, March 2023

    work page arXiv 2023

  5. [5]

    Deep Distance Map Regression Network with Shape-aware Loss for Imbalanced Medical Image Segmentation.arXiv e-prints, page arXiv:2501.09116, January 2025

    Huiyu Li et al. Deep Distance Map Regression Network with Shape-aware Loss for Imbalanced Medical Image Segmentation.arXiv e-prints, page arXiv:2501.09116, January 2025

    work page arXiv 2025

  6. [6]

    Deep-learning-based decomposition of overlapping-sparse images: application at the vertex of simulated neutrino interactions.Commun

    Sa´ ul Alonso-Monsalve et al. Deep-learning-based decomposition of overlapping-sparse images: application at the vertex of simulated neutrino interactions.Commun. Phys., 7(1):173, 2024

    work page 2024

  7. [7]

    Abed Abud et al

    A. Abed Abud et al. Neutrino interaction vertex reconstruction in DUNE with Pandora deep learning.Eur. Phys. J. C, 85(697):697, 2025

    work page 2025

  8. [8]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. InMedical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–241. Springer, 2015

    work page 2015

  9. [9]

    Mask R-CNN

    Kaiming He, Georgia Gkioxari, Piotr Doll´ ar, and Ross Girshick. Mask R-CNN.arXiv e-prints, page arXiv:1703.06870, March 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  10. [10]

    Dixon-VIBE Deep Learning (DIVIDE) Pseudo-CT Synthesis for Pelvis PET/MR Attenuation Correction.Journal of Nuclear Medicine, 2018

    Angel Torrado-Carvajal et al. Dixon-VIBE Deep Learning (DIVIDE) Pseudo-CT Synthesis for Pelvis PET/MR Attenuation Correction.Journal of Nuclear Medicine, 2018

    work page 2018

  11. [11]

    Burke et al

    Colin J. Burke et al. Deblending and classifying astronomical sources with Mask R-CNN deep learning.MNRAS, 490(3):3952–3965, December 2019. 12 REFERENCESMIGDAL

    work page 2019

  12. [12]

    Deblending overlapping galaxies in DECaLS using transformer-based algorithm: A method combining multiple bands and data types.PASA, 41:e035, May 2024

    Ran Zhang et al. Deblending overlapping galaxies in DECaLS using transformer-based algorithm: A method combining multiple bands and data types.PASA, 41:e035, May 2024

    work page 2024

  13. [13]

    Walter, Sebastian Damrich, and Fred A

    Florin C. Walter, Sebastian Damrich, and Fred A. Hamprecht. MultiStar: Instance Segmentation of Overlapping Objects with Star-Convex Polygons.arXiv e-prints, page arXiv:2011.13228, November 2020

    work page arXiv 2011

  14. [14]

    Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers.arXiv e-prints, page arXiv:2103.12340, March 2021

    Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers.arXiv e-prints, page arXiv:2103.12340, March 2021

    work page arXiv 2021

  15. [15]

    Alvarez et al

    V. Alvarez et al. NEXT-100 Technical Design Report (TDR): Executive Summary.JINST, 7:T06001, 2012

    work page 2012

  16. [16]

    Kekic et al

    M. Kekic et al. Demonstration of background rejection using deep convolutional neural networks in the NEXT experiment.JHEP, 01:189, 2021

    work page 2021

  17. [17]

    Sim´ on et al

    A. Sim´ on et al. Boosting background suppression in the NEXT experiment through Richardson-Lucy deconvolution.JHEP, 07:146, 2021

    work page 2021

  18. [18]

    Miernik et al

    K. Miernik et al. Optical Time Projection Chamber for imaging nuclear decays.Nuclear Instruments and Methods in Physics Research, Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 581(1-2):194–197, oct 2007

    work page 2007

  19. [19]

    Miernik, K. et al. Two-proton radioactivity of 45Fe.Eur. Phys. J. A, 42(3):431–439, 2009

    work page 2009

  20. [20]

    Pomorski et al

    M. Pomorski et al. Proton spectroscopy of 48Ni,46 Fe, and 44Cr.Phys. Rev. C, 90:014311, Jul 2014

    work page 2014

  21. [21]

    Acciarri et al

    R. Acciarri et al. Design and Construction of the MicroBooNE Detector.JINST, 12(02):P02017, 2017

    work page 2017

  22. [22]

    Abratenko et al

    P. Abratenko et al. Semantic segmentation with a sparse convolutional neural network for event reconstruction in MicroBooNE.Phys. Rev. D, 103(5):052012, 2021

    work page 2021

  23. [23]

    Abratenko et al

    P. Abratenko et al. Convolutional neural network for multiple particle identification in the MicroBooNE liquid argon time projection chamber.Phys. Rev. D, 103(9):092003, 2021

    work page 2021

  24. [24]

    Abratenko et al

    P. Abratenko et al. Search for an anomalous excess of charged-current quasielasticνe interactions with the MicroBooNE experiment using Deep-Learning-based reconstruction. Phys. Rev. D, 105(11):112003, 2022

    work page 2022

  25. [25]

    Grizzi, Margaret Voetberg, Giuseppe Cerati, Hadi Meidani, and V

    Vitor F. Grizzi, Margaret Voetberg, Giuseppe Cerati, Hadi Meidani, and V. Hewes. NuGraph2 with Context-Aware Inputs: Physics-Inspired Improvements in Semantic Segmentation. 9 2025

    work page 2025

  26. [26]

    Aurisano, V

    A. Aurisano, V. Hewes, G. Cerati, J. Kowalkowski, C. S. Lee, W. Liao, D. Grzenda, K. Gumpula, and X. Zhang. Graph neural network for neutrino physics event reconstruction. Phys. Rev. D, 110(3):032008, 2024

    work page 2024

  27. [27]

    H. M. Ara´ ujo et al. The MIGDAL experiment: Measuring a rare atomic process to aid the search for dark matter.Astropart. Phys., 151:102853, 2023

    work page 2023

  28. [28]

    Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning.arXiv e-prints, page arXiv:1702.03118, February 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  29. [29]

    Total Variation Applications in Computer Vision

    Vania V. Estrela, Hermes Aguiar Magalhaes, and Osamu Saotome. Total Variation Applications in Computer Vision.arXiv e-prints, page arXiv:1603.09599, March 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  30. [30]

    A. Migdal. Ionizatsiya atomov pri yadernykh reaktsiyakh (Ionization of atoms in nuclear reactions).ZhETF, 9:1163–1165, 1939

    work page 1939

  31. [31]

    Migdal Effect in Dark Matter Direct Detection Experiments.JHEP, 03:194, 2018

    Masahiro Ibe, Wakutaka Nakano, Yutaro Shoji, and Kazumine Suzuki. Migdal Effect in Dark Matter Direct Detection Experiments.JHEP, 03:194, 2018

    work page 2018

  32. [32]

    D. S. Akerib et al. Results of a Search for Sub-GeV Dark Matter Using 2013 LUX Data.Phys. Rev. Lett., 122(13):131301, 2019. 13 REFERENCESMIGDAL

    work page 2013

  33. [33]

    Armengaud et al

    E. Armengaud et al. Searching for low-mass dark matter particles with a massive Ge bolometer operated above-ground.Phys. Rev. D, 99(8):082003, 2019

    work page 2019

  34. [34]

    Z. Z. Liu et al. Constraints on Spin-Independent Nucleus Scattering with sub-GeV Weakly Interacting Massive Particle Dark Matter from the CDEX-1B Experiment at the China Jinping Underground Laboratory.Phys. Rev. Lett., 123(16):161301, 2019

    work page 2019

  35. [35]

    Aprile et al

    E. Aprile et al. Search for Light Dark Matter Interactions Enhanced by the Migdal Effect or Bremsstrahlung in XENON1T.Phys. Rev. Lett., 123(24):241803, 2019

    work page 2019

  36. [36]

    SENSEI: Direct-Detection Results on sub-GeV Dark Matter from a New Skipper-CCD.Phys

    Liron Barak et al. SENSEI: Direct-Detection Results on sub-GeV Dark Matter from a New Skipper-CCD.Phys. Rev. Lett., 125(17):171802, 2020

    work page 2020

  37. [37]

    Adhikari et al

    G. Adhikari et al. Searching for low-mass dark matter via the Migdal effect in COSINE-100. Phys. Rev. D, 105(4):042006, 2022

    work page 2022

  38. [38]

    Armengaud et al

    E. Armengaud et al. Search for sub-GeV dark matter via the Migdal effect with an EDELWEISS germanium detector with NbSi transition-edge sensors.Phys. Rev. D, 106(6):062004, 2022

    work page 2022

  39. [39]

    Agnes et al

    P. Agnes et al. Search for Dark-Matter–Nucleon Interactions via Migdal Effect with DarkSide-50.Phys. Rev. Lett., 130(10):101001, 2023

    work page 2023

  40. [40]

    Abe et al

    K. Abe et al. Direct dark matter searches with the full data set of XMASS-I.Phys. Rev. D, 108(8):083022, 2023

    work page 2023

  41. [41]

    Aalbers et al

    J. Aalbers et al. Search for new physics in low-energy electron recoils from the first LZ exposure.Phys. Rev. D, 108(7):072006, 2023

    work page 2023

  42. [42]

    Search for Dark-Matter–Nucleon Interactions with a Dark Mediator in PandaX-4T.Phys

    Di Huang et al. Search for Dark-Matter–Nucleon Interactions with a Dark Mediator in PandaX-4T.Phys. Rev. Lett., 131(19):191002, 2023

    work page 2023

  43. [43]

    M. F. Albakry et al. Search for low-mass dark matter via bremsstrahlung radiation and the Migdal effect in SuperCDMS.Phys. Rev. D, 107(11):112013, 2023

    work page 2023

  44. [44]

    First Direct-Detection Results on Sub-GeV Dark Matter Using the SENSEI Detector at SNOLAB.Phys

    Prakruth Adari et al. First Direct-Detection Results on Sub-GeV Dark Matter Using the SENSEI Detector at SNOLAB.Phys. Rev. Lett., 134(1):011804, 2025

    work page 2025

  45. [45]

    Aggarwal et al

    K. Aggarwal et al. Probing Benchmark Models of Hidden-Sector Dark Matter with DAMIC-M.Phys. Rev. Lett., 135(7):071002, 2025

    work page 2025

  46. [46]

    Search for the Migdal effect in liquid xenon with keV-level nuclear recoils

    Jingke Xu et al. Search for the Migdal effect in liquid xenon with keV-level nuclear recoils. Phys. Rev. D, 109(5):L051101, 2024

    work page 2024

  47. [47]

    Technical report, Hamamatsu Photonics, 05 2022

    Orca-quest qcmos camera c15550-20up technical note. Technical report, Hamamatsu Photonics, 05 2022

    work page 2022

  48. [48]

    Performance of the Optical Time Projection Chamber of the MIGDAL Experiment.[In Progress], 2025

    MIGDAL Collaboration. Performance of the Optical Time Projection Chamber of the MIGDAL Experiment.[In Progress], 2025

    work page 2025

  49. [49]

    Schueler et al

    J. Schueler et al. Transforming a rare event search into a not-so-rare event search in real-time with deep learning-based object detection.Phys. Rev. D, 111(7):072004, 2025

    work page 2025

  50. [50]

    Dolan, Christopher McCabe, and Harry M

    Peter Cox, Matthew J. Dolan, Christopher McCabe, and Harry M. Quiney. Precise predictions and new insights for atomic ionization from the Migdal effect.Phys. Rev. D, 107(3):035032, 2023

    work page 2023

  51. [51]

    An Unbiased Detector of Curvilinear Structures.IEEE Trans

    Carsten Steger. An Unbiased Detector of Curvilinear Structures.IEEE Trans. Pattern Anal. Mach. Intell., 20:113–125, 1998

    work page 1998

  52. [52]

    3D track reconstruction of low-energy electrons in the MIGDAL low pressure optical time projection chamber.JINST, 18(07):C07013, 2023

    Elizabeth Tilly and Magnus Handley. 3D track reconstruction of low-energy electrons in the MIGDAL low pressure optical time projection chamber.JINST, 18(07):C07013, 2023

    work page 2023

  53. [53]

    Ultralytics YOLOv8, 2023

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics YOLOv8, 2023

    work page 2023

  54. [54]

    Sim2real in robotics and automation: Applications and challenges

    Sebastian H¨ ofer et al. Sim2real in robotics and automation: Applications and challenges. IEEE Transactions on Automation Science and Engineering, 18(2):398–400, 2021. 14 REFERENCESMIGDAL

    work page 2021

  55. [55]

    Edgeman et al

    D. Edgeman et al. Optical effects in Gaseous Electron Multipliers.[In Progress], 2025

    work page 2025

  56. [56]

    Available at https://doi.org/10.5281/zenodo.12628437andhttps://github.com/jschuel/migYOLO

    Jeffrey Schueler.migYOLO, July 2024. Available at https://doi.org/10.5281/zenodo.12628437andhttps://github.com/jschuel/migYOLO

    work page doi:10.5281/zenodo.12628437andhttps://github.com/jschuel/migyolo 2024

  57. [57]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization.arXiv e-prints, page arXiv:1711.05101, November 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  58. [58]

    Springer Berlin Heidelberg, Berlin, Heidelberg, 2012

    Lutz Prechelt.Early Stopping — But When?, pages 53–67. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012

    work page 2012

  59. [59]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

    work page 2019

  60. [60]

    Galaxy Zoo DECaLS: Detailed visual morphology measurements from volunteers and deep learning for 314 000 galaxies.MNRAS, 509(3):3966–3988, January 2022

    Mike Walmsley et al. Galaxy Zoo DECaLS: Detailed visual morphology measurements from volunteers and deep learning for 314 000 galaxies.MNRAS, 509(3):3966–3988, January 2022

    work page 2022

  61. [61]

    Principal curves.Journal of the American Statistical Association, 84(406):502–516, June 1989

    Trevor Hastie and Werner Stuetzle. Principal curves.Journal of the American Statistical Association, 84(406):502–516, June 1989. https://web.stanford.edu/~hastie/Papers/principalcurves.pdf

    work page 1989

  62. [62]

    princurve 2.0: Fit a principal curve in arbitrary dimension, June 2018

    Robrecht Cannoodt. princurve 2.0: Fit a principal curve in arbitrary dimension, June 2018

    work page 2018

  63. [63]

    R Foundation for Statistical Computing, Vienna, Austria, 2023

    R Core Team.R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2023. 15

    work page 2023