Pith. sign in

REVIEW 3 major objections 5 minor 67 references

Maybe you don't need a U-Net: convolutional feature upsampling for materials micrograph segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A once-trained convolutional upsampler lifts coarse DINOv2 features to pixel resolution, letting interactive classifiers segment micrographs faster and with fewer labels than fine-tuning a U-Net.

desk verdict Solid engineering paper with an honest but underspecified efficiency claim: the per-image PCA preprocessing likely makes the 'under 10s' timing numbers optimistic; still worth citing and reviewing. read the letter →

arxiv 2508.21529 v1 pith:66UE3O23 submitted 2025-08-29 cs.CV cond-mat.mtrl-sci

classification cs.CVcond-mat.mtrl-sci
keywords featureupsamplingDINOv2interactivesegmentationmicrographmaterialssciencefoundationmodelssparselabelsconvolutionalneuralnetwork
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

Foundation-model vision features describe image patches, not pixels, which limits their use on microscopy images where fine phases, hairline cracks, and large fields of view are common. This paper trains a small convolutional network to upsample low-resolution DINOv2 features to full resolution, using high-resolution features produced by a per-image FeatUp implicit model as training targets. At test time the upsampler runs without retraining, and its features are combined with classical pixel features before an XGBoost classifier maps sparse user labels to a full segmentation. The central claim is that this workflow yields accurate micrograph segmentations in seconds and with far fewer labels than training or fine-tuning a U-Net such as MicroNet.

What carries the argument

The engine of the method is a convolutional upsampler trained once on natural images: a learned downsampler extracts image guidance at multiple scales, and a U-Net-style upsampler takes the low-resolution DINOv2 features, with a shared PCA applied, and progressively doubles resolution to full pixel level under supervision from the high-resolution features of a FeatUp implicit model via smooth L1 loss. At inference the same upsampler is applied to new micrographs without retraining, producing a compact k-channel feature stack (k from 16 to 128) that is concatenated to classical local-intensity, edge, and texture features and classified by XGBoost. The shared PCA computed over 50 transformed views is load-bearing: it aligns the distribution of low-resolution features with the high-resolution training targets, and it also makes compressed upsamplers possible by keeping only the first k principal components.

What would settle it

Measure the end-to-end wall-clock time to featurise a roughly 2000 by 1000 pixel micrograph on a 4 GB laptop GPU with the full preprocessing pipeline including the shared PCA, and compare it with the claimed under-10-second figure; a second check is to run the benchmark with the number of PCA transforms set to 1 instead of 50 and see whether the mIoU gains survive.

Watch

Extended reading notes

Core claim

The paper's central claim is that upsampled deep features, not a task-specific segmentation network, are the missing ingredient for practical micrograph segmentation. On three benchmark datasets (nickel superalloys, T-cells, copper ore), replacing the classical feature stack alone with classical features plus upsampled DINOv2 features raises class-averaged mIoU, for example from 0.56 to 0.75 on the Ni superalloy dataset with only four labelled images. With sparse labels the approach reaches 75% mIoU in about 20 seconds of featurisation, classifier training, and application, whereas the MicroNet U-Net takes around four minutes to reach 80% mIoU with full labels and trains poorly on sparse labels. The paper also claims that the upsampler generalises across instruments (SEM, TEM, reflected-light microscopy) without ever being trained on micrographs.

Load-bearing premise

The method's stated speed and accuracy depend on an inference-time preprocessing step, a shared PCA computed over transformed versions of the input before upsampling, and the paper does not say how many transforms are used at inference or how that cost enters the reported timings.

Editorial extensions

If this is right

  • Interactive segmentation with upsampled deep features reaches high mIoU much earlier in wall-clock time than fine-tuning a U-Net, so a microscopist can correct mistakes while the CNN would still be training.
  • The frozen upsampler transfers to unseen material systems and imaging modalities without additional training, as long as classical features and user labels are supplied.
  • Compressed feature stacks make large micrographs around 2000 by 1000 pixels treatable on a 4 GB laptop GPU and make storage of multi-image datasets cheaper.
  • Given enough time and fully labelled data, a fine-tuned task-specific CNN is expected to eventually outperform the interactive approach, so the advantage is in speed and label efficiency rather than ceiling accuracy.
  • Upsampling sharpens segmentation of sub-patch features such as cracks, an advantage that becomes more important as vision foundation models move to larger patch sizes.

Reading between the lines

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

  • A natural stress test is to vary the number of PCA transforms at inference: if one or a few transforms suffice, the speed claim holds directly, but if all 50 are required, the reported timings need to be revisited.
  • Because the upsampler is trained on RGB natural images, retraining it on micrograph data or on a larger generalist dataset is a direct extension the authors flag as future work.
  • The feature extraction is one-time and task-agnostic, so the same upsampled deep features could plausibly support denoising, property prediction, or other dense image tasks in materials imaging without retraining.
  • An interactive benefit the paper notes but does not benchmark as a headline result is rapid relabelling: after the first featurisation, users can add labels and retrain only the classifier, making iterative segmentation corrections cheap.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper introduces a lightweight convolutional upsampler that maps low-resolution DINOv2 patch features to high-resolution features, trained to regress to FeatUp implicit ground-truth features. The upsampled features are concatenated with classical hand-crafted features and fed to an XGBoost classifier for interactive segmentation of micrographs. The authors report improved mIoU over classical features alone on three materials datasets, a qualitative and quantitative speed advantage over fine-tuning a MicroNet U-Net, and control experiments showing the improvement is not due to added channels. The central claim is that interactive segmentation with these deep features is "far faster and with far fewer labels" than training or fine-tuning a traditional convolutional network.

Significance. If the timing and accuracy claims hold, this is a practically useful contribution: it offers a memory- and time-efficient way to use foundation-model features for interactive materials-image segmentation on a laptop GPU, with reproducible code and public data. The control experiments in Section S7, the multi-dataset benchmark, and the explicit comparison to a U-Net baseline are valuable. The main risks are (i) the per-image PCA preprocessing is not accounted for in the reported timings, which may undermine the central efficiency claim, and (ii) the evaluation metric has a definitional error and the benchmark protocol includes training images in the reported mIoU. Both are fixable and should be addressed before publication.

major comments (3)
  1. [§3.2, §S3, Figure 4, Figure 7] The paper states that the FeatUp-compatible preprocessing—a shared PCA over Nt=50 transformations of the input—"must therefore be applied during inference" (Section S3), and Figure 4 claims "under 10s on a laptop GPU" while Figure 7 reports 20.4s for featurisation + classifier training + application over 22 images. However, the paper never states how many transformations are used at inference or whether the reported timings include the cost of running DINOv2 on those transformations and fitting/applying the PCA. If Nt=50 is required, each 2000x1000 micrograph would need roughly 50 DINOv2-S forward passes plus a PCA over about 5e5 patch vectors, which is plausibly minutes on a 4GB RTX 3050M, contradicting the stated speed. If a smaller Nt is used, the upsampler input is distributionally shifted from its training input, as Figure S2b shows even N=2-3 changes the subspace substantially. The authors should state the inference-time Nt, include its full cost in all reported timings, and either demonstrate that a small Nt preserves quality or revise the speed claims accordingly.
  2. [§3.4] The definition of mIoU is incorrect: the text says it "measures the ratio of true positives to true positives plus false positives (i.e, the positive predictive value)", which is precision, not Intersection over Union. IoU is TP/(TP+FP+FN). Please correct the definition and explicitly confirm that all reported mIoU values in Table 1, Table S2, Figure 6, and Figure 7 were computed with the standard formula; if precision was used instead, the results must be recomputed.
  3. [§4.1, Table 1, Figure 7] The benchmark protocol says classifiers are trained on four images and "applied unseen to the rest" of the dataset, but the reported mIoU is "measured across all images (train included)". Evaluating on training images inflates the scores and weakens the generalization claim made in the text. Please report held-out mIoU separately, or at least provide both train and held-out numbers for every configuration in Table 1 and Figure 7, so readers can assess generalization and the comparison to MicroNet on equal terms.
minor comments (5)
  1. [References] Reference [8] duplicates reference [2], and reference [53] duplicates reference [4]; please consolidate these duplicate entries.
  2. [§S5, Table S2] Table S2 reports class-avg mIoU 0.53±0.14 for the Ni dataset with 4 labelled examples and 0 added channel channels, whereas Table 1 reports 0.56±0.16 for the same configuration; clarify whether these were obtained under the same protocol or explain the difference.
  3. [§3.1, Table S1] The loss is referred to as "smooth MSE" in Section 3.1 but as "Smooth L1" in Table S1; please use one name consistently.
  4. [Figure 3 caption] The caption states "bottom-left is better" without specifying which axes correspond to time and memory and which corresponds to feature quality; please make the axes explicit so the statement is interpretable.
  5. [Figure 2 caption] There is a typo: "Implict" should be "Implicit" in the caption text.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the upsampler is supervised by external FeatUp targets and accuracy is measured against independent ground-truth labels.

full rationale

The paper's central derivation trains a convolutional upsampler to map low-resolution DINOv2 features, plus image guidance, to high-resolution features produced by FeatUp's per-image implicit network. Section 3.1 states: 'Our approach is to learn a convolutional upsampler network to map from an image and its corresponding low-resolution features as extracted by a ViT to the high-resolution map produced by training an implicit FeatUp network on that image.' This is a supervised regression to an external method's outputs, not a quantity fitted from the segmentation labels used in evaluation. The segmentation experiments use independent ground truths from Ni superalloys, human T-cells, and copper ore, and the same labels and classifier are used for both feature-sets. Section S7 explicitly tests whether the gain comes from mere extra channels, showing that random noise, uniform zeros, and duplicated classical channels do not reproduce the +HR ViT improvement. The disclosed distribution-shift issue in Section S3 and the unstated number of inference-time PCA transforms are reproducibility and timing concerns, not circularity: the PCA preprocessing is applied consistently in training and inference, and the final claim is still settled by external labels. Self-citations to the authors' prior work [24] appear as methodological continuity ('Following our previous work [24], we concatenate these upsampled deep features with the classical features'), not as the proof of the central efficiency or accuracy claims, and the comparison against MicroNet is an external empirical benchmark. No step in the derivation reduces by construction to its own inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim relies on a small set of user-chosen parameters (PCA dimension, number of transforms, training set size) and on several domain assumptions about the validity of FeatUp targets, DINOv2 transfer, and the benchmark labels. No new physical entities are introduced.

free parameters (4)
  • PCA components k = 128 (default; 16/32/64 compressed variants tested)
    Number of upsampled feature channels; chosen by the authors, affects memory, speed and segmentation quality (Table S2 shows mIoU saturates near 32-128).
  • Transforms Nt for FeatUp PCA = 50 (approximation of 3000)
    Number of image transformations used to compute the shared PCA that preprocesses LR features; affects the feature distribution and the test-time computational cost (Section 3.2 and S3).
  • Training set size = 2000 ImageNet Reduced images
    Subset used to train the upsampler; chosen for compute budget; authors state larger sets could improve performance (Section 3.1).
  • Upsampler hyperparameters = lr=1e-4, batch=32, 5000 epochs, smooth L1
    Standard training choices from Table S1; not fitted to downstream segmentation targets.
assumptions (5)
  • domain assumption FeatUp implicit high-resolution features are a valid training target for the upsampler.
    The paper trains the CNN to regress to FeatUp implicit outputs without independently validating that those outputs are better than cheaper targets (Section 3.1).
  • domain assumption DINOv2 features pretrained on natural images transfer to micrographs.
    The whole approach depends on this; the authors acknowledge the RGB/greyscale domain gap as a limitation (Section 5).
  • ad hoc to paper A PCA over 50 transformations approximates the full FeatUp PCA distribution.
    Stated in Section 3.2 and S3; this choice defines the preprocessing and its distribution shift.
  • domain assumption Classical features combined with HR ViT features retain all useful information of the classical stack.
    The concatenation strategy assumes no negative interference; the control experiments in S7 partially support this.
  • domain assumption Ground-truth segmentations in the three benchmark datasets are correct.
    Ni, T-cell and Cu ore labels are taken from external sources (refs 7, 50, 51) and used to compute mIoU.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Maybe you don't need a U-Net: convolutional feature upsampling for materials micrograph segmentation." pith.science (2026). https://pith.science/paper/66UE3O23

@misc{pith2026250821529,
  author       = {Pith},
  title        = {Pith review of: Maybe you don't need a U-Net: convolutional feature upsampling for materials micrograph segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66UE3O23}},
  note         = {Machine review of arXiv:2508.21529}
}
read the original abstract

Feature foundation models - usually vision transformers - offer rich semantic descriptors of images, useful for downstream tasks such as (interactive) segmentation and object detection. For computational efficiency these descriptors are often patch-based, and so struggle to represent the fine features often present in micrographs; they also struggle with the large image sizes present in materials and biological image analysis. In this work, we train a convolutional neural network to upsample low-resolution (i.e, large patch size) foundation model features with reference to the input image. We apply this upsampler network (without any further training) to efficiently featurise and then segment a variety of microscopy images, including plant cells, a lithium-ion battery cathode and organic crystals. The richness of these upsampled features admits separation of hard to segment phases, like hairline cracks. We demonstrate that interactive segmentation with these deep features produces high-quality segmentations far faster and with far fewer labels than training or finetuning a more traditional convolutional network.

Figures

Figures reproduced from arXiv: 2508.21529 by the authors.

Figure 1
Figure 1. An explanation of our main contributions: we ef￾ficiently upsample coarse semantic features from DINOv228 and combine it with pixel level classical features for interac￾tive segmentation. The richness of the ViT features allows the classifier (trained to map from features → user labels) to perform more accurate segmentation of the blocky α− and needle-like β−polymorph phases of an XCT cross-section of glutamic acid3… view at source ↗
Figure 2
Figure 2. a. Model diagram for our upsampler. The input image xi has its low-resolution (LR) features xf computed from the ViT and pre-processed into a ‘FeatUp compatible’ form. It is downsampled log2(p) times via a learned convolutional downsampler, whose activations are used as guidance for the convolutional feature upsampler. The result is compared to the FeatUp Implict high-resolution (HR) ‘ground truth’ via a smooth MSE … view at source ↗
Figure 3
Figure 3. Performance landscape of various ViT feature upsamplers. a. the input image and corresponding patch-based feature vectors extracted by DINOv2. All feature vectors have been reduced to 3 dimensions via PCA and plotted as RGB channels. b. memory, time and feature-quality comparison for input image at size (336, 336); bottom-left is better. c. time and memory scaling as image side length is increased. ‘LiFT’ uses DINO … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Interactive segmentation results on SEM micrographs using ‘Classical’ features and ‘Classical + HR ViT fea￾tures’, which are features from DINOv2-S which have been upsampled using our model. The labels and classifier type were the same across each example. The addition…
Figure 5
Figure 5. Figure 5: Generalisation results of classifiers trained on classical and +HR ViT features over three datasets: SEM images of Nickel superalloys, TEM images of human T-cells and reflected light microscopy of copper ore in resin. The classifiers are trained on labels and features …
Figure 6
Figure 6. Figure 6: Class-averaged mIoU as function of number of labelled images for the various benchmark datasets for each feature-set. We see the +HR ViT features outperform the classical feature-set for each dataset. Example predictions for these ‘fully-trained’ classi￾fiers on the Ni…
Figure 7
Figure 7. Figure 7: Comparison of segmentation performance for a CNN vs. interactive segmentation supplemented with upsampled deep features (HR ViT) on the 3-phase Ni superalloy dataset. Four micrographs were used for training, either with ‘sparse’ labels from Section 4.1 or ‘full’ labels…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 47 canonical work pages

  1. [1]

    Artificial neural network approach for multi- phase segmentation of battery electrode nano-CT images,

    Z. Su, E. Decenci `ere, T.-T. Nguyen, K. El-Amiry, V . De Andrade, A. A. Franco, and A. Demorti`ere, “Artificial neural network approach for multi- phase segmentation of battery electrode nano-CT images,” npj Computational Materials , vol. 8, p. 30, Feb. 2022

  2. [3]

    Methods—Kintsugi Imaging of Battery Electrodes: Distinguishing Pores from the Carbon Binder Domain using Pt Deposition,

    S. J. Cooper, S. A. Roberts, Z. Liu, and B. Winiarski, “Methods—Kintsugi Imaging of Battery Electrodes: Distinguishing Pores from the Carbon Binder Domain using Pt Deposition,” Journal of The Electrochemical Society, vol. 169, p. 070512, July 2022

  3. [5]

    Au- tomated segmentation of large image datasets using artificial intelligence for microstructure characterisation and damage analysis,

    S. Medghalchi, J. Kortmann, S.-H. Lee, E. Karimi, U. Kerzel, and S. Korte-Kerzel, “Au- tomated segmentation of large image datasets using artificial intelligence for microstructure characterisation and damage analysis,”Materials & Design, vol. 243, p. 113031, 2024

  4. [6]

    Guiding the Design of Heteroge- neous Electrode Microstructures for Li-Ion Bat- teries: Microscopic Imaging, Predictive Model- ing, and Machine Learning,

    H. Xu, J. Zhu, D. P. Finegan, H. Zhao, X. Lu, W. Li, N. Ho ffman, A. Bertei, P. Shearing, and M. Z. Bazant, “Guiding the Design of Heteroge- neous Electrode Microstructures for Li-Ion Bat- teries: Microscopic Imaging, Predictive Model- ing, and Machine Learning,” Advanced Energy Materials, vol. 11, p. 2003908, May 2021. Pub- lisher: John Wiley & Sons, Ltd

  5. [7]

    Microstructure segmentation with deep learn- ing encoders pre-trained on a large microscopy dataset,

    J. Stuckner, B. Harder, and T. M. Smith, “Microstructure segmentation with deep learn- ing encoders pre-trained on a large microscopy dataset,” npj Computational Materials , vol. 8, p. 200, Sept. 2022

  6. [8]

    Quantification and modeling of mechanical degradation in lithium-ion batteries based on nanoscale imaging,

    S. M ¨uller, P. Pietsch, B.-E. Brandt, P. Baade, V . De Andrade, F. De Carlo, and V . Wood, “Quantification and modeling of mechanical degradation in lithium-ion batteries based on nanoscale imaging,” Nature Communications , vol. 9, p. 2340, June 2018

  7. [9]

    TauFactor: An open-source appli- 11 Docherty et al. Feature Upsampling & Micrograph Segmentation Preprint cation for calculating tortuosity factors from to- mographic data,

    S. Cooper, A. Bertei, P. Shearing, J. Kilner, and N. Brandon, “TauFactor: An open-source appli- 11 Docherty et al. Feature Upsampling & Micrograph Segmentation Preprint cation for calculating tortuosity factors from to- mographic data,” SoftwareX, vol. 5, pp. 203–210, 2016

  8. [10]

    Taufactor 2: A gpu accelerated python tool for microstruc- tural analysis,

    S. Kench, I. Squires, and S. Cooper, “Taufactor 2: A gpu accelerated python tool for microstruc- tural analysis,” Journal of Open Source Software, vol. 8, no. 88, p. 5358, 2023

Show all 67 references
  1. [11]

    X-ray computed tomography,

    P. J. Withers, C. Bouman, S. Carmignato, V . Cnudde, D. Grimaldi, C. K. Hagen, E. Maire, M. Manley, A. Du Plessis, and S. R. Stock, “X-ray computed tomography,” Nature Reviews Methods Primers, vol. 1, p. 18, Feb. 2021

  2. [12]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” 2015

  3. [13]

    Graph-constrained Contrastive Regularization for Semi-weakly V olumetric Seg- mentation,

    S. Reiß, C. Seibold, A. Freytag, E. Rodner, and R. Stiefelhagen, “Graph-constrained Contrastive Regularization for Semi-weakly V olumetric Seg- mentation,”

  4. [14]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. White- head, A. C. Berg, W.-Y . Lo, P. Doll ´ar, and R. Girshick, “Segment anything,”arXiv preprint, arXiv:2304.02643, 2023

  5. [15]

    SAMBA: A Trainable Segmenta- tion Web-App with Smart Labelling,

    R. Docherty, I. Squires, A. Vamvakeros, and S. J. Cooper, “SAMBA: A Trainable Segmenta- tion Web-App with Smart Labelling,” Journal of Open Source Software , vol. 9, no. 98, p. 6159,

  6. [16]

    Cellpose 2.0: how to train your own model,

    M. Pachitariu and C. Stringer, “Cellpose 2.0: how to train your own model,” Nature Methods, Nov. 2022

  7. [17]

    A Threshold Selection Method from Gray-Level Histograms,

    N. Otsu, “A Threshold Selection Method from Gray-Level Histograms,” IEEE Transactions on Systems, Man, and Cybernetics , vol. 9, no. 1, pp. 62–66, 1979

  8. [18]

    Algorithm as 136: A k-means clustering algorithm,

    J. A. Hartigan and M. A. Wong, “Algorithm as 136: A k-means clustering algorithm,” Journal of the Royal Statistical Society. Series C (Applied Statistics), vol. 28, no. 1, pp. 100–108, 1979

  9. [19]

    Topographic distance and watershed lines,

    F. Meyer, “Topographic distance and watershed lines,” Signal Processing, vol. 38, no. 1, pp. 113– 125, 1994

  10. [20]

    A Generalization of Otsu’s Method and Minimum Error Thresholding,

    J. T. Barron, “A Generalization of Otsu’s Method and Minimum Error Thresholding,” arXiv preprint, arXiv:2007.07350, 2020

  11. [21]

    Resolving the Discrep- ancy in Tortuosity Factor Estimation for Li-Ion Battery Electrodes through Micro-Macro Mod- eling and Experiment,

    F. L. E. Usseglio-Viretta, A. Colclasure, A. N. Mistry, K. P. Y . Claver, F. Pouraghajan, D. P. Finegan, T. M. M. Heenan, D. Abraham, P. P. Mukherjee, D. Wheeler, P. Shearing, S. J. Cooper, and K. Smith, “Resolving the Discrep- ancy in Tortuosity Factor Estimation for Li-Ion B...

  12. [22]

    Trainable Weka Segmen- tation: a machine learning tool for microscopy pixel classification,

    I. Arganda-Carreras, V . Kaynig, C. Rueden, K. W. Eliceiri, J. Schindelin, A. Cardona, and H. Sebastian Seung, “Trainable Weka Segmen- tation: a machine learning tool for microscopy pixel classification,” Bioinformatics, vol. 33, pp. 2424–2426, Aug. 2017

  13. [23]

    ilastik: interactive machine learn- ing for (bio)image analysis,

    S. Berg, D. Kutra, T. Kroeger, C. N. Straehle, B. X. Kausler, C. Haubold, M. Schiegg, J. Ales, T. Beier, M. Rudy, K. Eren, J. I. Cer- vantes, B. Xu, F. Beuttenmueller, A. Wolny, C. Zhang, U. Koethe, F. A. Hamprecht, and A. Kreshuk, “ilastik: interactive machine learn- ing for ...

  14. [24]

    Upsampling dinov2 features for un- supervised vision tasks and weakly super- vised materials segmentation,

    R. Docherty, A. Vamvakeros, and S. J. Cooper, “Upsampling dinov2 features for un- supervised vision tasks and weakly super- vised materials segmentation,” arXiv preprint, arXiv:2410.19836, 2024

  15. [25]

    Auto-Context and Its Ap- plication to High-Level Vision Tasks and 3D Brain Image Segmentation,

    X. Bai and Z. Tu, “Auto-Context and Its Ap- plication to High-Level Vision Tasks and 3D Brain Image Segmentation,” IEEE Transactions on Pattern Analysis & Machine Intelli- gence, vol. 32, pp. 1744–1757, Oct. 2010. Place: Los Alamitos, CA, USA Publisher: IEEE Com- puter Society

  16. [26]

    ExpertSegmentation: Segmentation for mi- croscopy with domain-informed targets via cus- tom loss,

    N. Prakash, P. Gasper, and F. Usseglio-Viretta, “ExpertSegmentation: Segmentation for mi- croscopy with domain-informed targets via cus- tom loss,” Acta Materialia, vol. 291, p. 120993, 2025

  17. [27]

    Emerg- ing Properties in Self-Supervised Vision Trans- formers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerg- ing Properties in Self-Supervised Vision Trans- formers,” arXiv preprint, arXiv:2104.14294 , 2021

  18. [28]

    DINOv2: Learn- ing Robust Visual Features without Supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.- W. Li, I. Misra, M. Rabbat, V . Sharma, G. Syn- naeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, ...

  19. [29]

    Self-supervised learning from images with a joint-embedding predictive architecture,

    M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y . LeCun, and N. Bal- las, “Self-supervised learning from images with a joint-embedding predictive architecture,”arXiv preprint, arXiv:2301.08243, 2023. 12 Docherty et al. Feature Upsampling & Micrograph Seg...

  20. [30]

    Deep ViT Features as Dense Visual Descrip- tors,

    S. Amir, Y . Gandelsman, S. Bagon, and T. Dekel, “Deep ViT Features as Dense Visual Descrip- tors,” arXiv preprint, arXiv:2112.05814, 2022

  21. [31]

    FeatUp: A Model-Agnostic Framework for Features at Any Resolution,

    S. Fu, M. Hamilton, L. Brandt, A. Feldman, Z. Zhang, and W. T. Freeman, “FeatUp: A Model-Agnostic Framework for Features at Any Resolution,” arXiv preprint, arXiv:2403.10516 , 2024

  22. [32]

    LiFT: A Surprisingly Simple Lightweight Feature Transform for Dense ViT Descriptors,

    S. Suri, M. Walmer, K. Gupta, and A. Shrivas- tava, “LiFT: A Surprisingly Simple Lightweight Feature Transform for Dense ViT Descriptors,” arXiv preprint, arXiv:2403.14625, 2024

  23. [33]

    LoftUp: Learning a Coordinate- Based Feature Upsampler for Vision Founda- tion Models,

    H. Huang, A. Chen, V . Havrylov, A. Geiger, and D. Zhang, “LoftUp: Learning a Coordinate- Based Feature Upsampler for Vision Founda- tion Models,” arXiv preprint, arXiv:2504.14032, 2025

  24. [34]

    haesleinhuepf/napari-accelerated-pixel-and- object- classification: 0.14.1,

    R. Haase, D. Lee, D. D. Pop, and L. ˇZigutyt˙e, “haesleinhuepf/napari-accelerated-pixel-and- object- classification: 0.14.1,” Nov. 2023

  25. [35]

    Measuring the Particle Packing of l-Glutamic Acid Crystals through X-ray Computed To- mography for Understanding Powder Flow and Consolidation Behavior,

    T. D. Turner, P. Gajjar, I. S. Fragkopoulos, J. Carr, T. T. H. Nguyen, D. Hooper, F. Clarke, N. Dawson, P. J. Withers, and K. J. Roberts, “Measuring the Particle Packing of l-Glutamic Acid Crystals through X-ray Computed To- mography for Understanding Powder Flow and Consolida...

  26. [36]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polo- sukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30,

  27. [37]

    An Image is Worth 16x16 Words: Transformers for Im- age Recognition at Scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. De- hghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Transformers for Im- age Recognition at Scale,” arXiv preprint, arXiv:2010.11929, 2021

  28. [38]

    Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Do- mains,

    M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. Barron, and R. Ng, “Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Do- mains,” arXiv preprint, arXiv:2006.10739, 2020

  29. [39]

    Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data,

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data,” arXiv preprint, arXiv:2401.10891, 2024

  30. [40]

    A cookbook of self-supervised learning,

    R. Balestriero, M. Ibrahim, V . Sobal, A. Morcos, S. Shekhar, T. Goldstein, F. Bordes, A. Bardes, G. Mialon, Y . Tian, A. Schwarzschild, A. G. Wilson, J. Geiping, Q. Garrido, P. Fernandez, A. Bar, H. Pirsiavash, Y . LeCun, and M. Gold- blum, “A cookbook of self-supervised lear...

  31. [41]

    Masked Autoencoders Are Scalable Vision Learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked Autoencoders Are Scalable Vision Learners,” arXiv preprint, arXiv:2111.06377, 2021

  32. [42]

    NeRF: Representing Scenes as Neural Radi- ance Fields for View Synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing Scenes as Neural Radi- ance Fields for View Synthesis,” arXiv preprint, arXiv:2003.08934, 2020

  33. [43]

    Vision Transformers Need Registers,

    T. Darcet, M. Oquab, J. Mairal, and P. Bo- janowski, “Vision Transformers Need Registers,” arXiv preprint, arXiv:2309.16588, 2023

  34. [44]

    Improving 2d feature represen- tations by 3d-aware fine-tuning,

    Y . Yue, A. Das, F. Engelmann, S. Tang, and J. E. Lenssen, “Improving 2d feature represen- tations by 3d-aware fine-tuning,” arXiv preprint, arXiv:2407.20229, 2024

  35. [45]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei- Fei, “Imagenet large scale visual recognition challenge,” arXiv preprint, arXiv:1409.0575 , 2015

  36. [46]

    Random Forests,

    L. Breiman, “Random Forests,” Machine Learn- ing, vol. 45, pp. 5–32, Oct. 2001

  37. [47]

    Xgboost: A scal- able tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scal- able tree boosting system,” in Proceedings of the 22nd ACM SIGKDD International Confer- ence on Knowledge Discovery and Data Mining , KDD ’16, p. 785–794, ACM, Aug. 2016

  38. [48]

    Biphase cathode sem

    “Biphase cathode sem.” https://forum.image.sc/t/ three-phase-segmentation-of-sem-image/ 92812

  39. [49]

    Electron micrographs

    B. D. University of Wisconsin Stevens Point, “Electron micrographs.” https://www4.uwsp. edu/biology/courses/botlab/Lab03c. htm

  40. [50]

    Semi-automatic deter- mination of cell surface areas used in systems biology.,

    V . Morath, M. Keuper, M. Rodriguez-Franco, S. Deswal, G. Fiala, B. Blumenthal, D. Kaschek, J. Timmer, G. Neuhaus, S. Ehl, O. Ronneberger, and W. W. A. Schamel, “Semi-automatic deter- mination of cell surface areas used in systems biology.,” Frontiers in bioscience (Elite edit...

  41. [51]

    M. P. Filippo, O. d. F. M. Gomes, G. A. O. P. d. Costa, and G. L. A. Mota, “Deep learn- ing semantic segmentation of opaque and non- opaque minerals from epoxy resin in reflected 13 Docherty et al. Feature Upsampling & Micrograph Segmentation Preprint light microscopy images,”...

  42. [52]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” arXiv preprint, arXiv:1512.03385, 2015

  43. [53]

    Utilizing active learning to accel- erate segmentation of microstructures with tiny annotation budgets,

    L. H. Rieger, F. Cadiou, Q. Jacquet, V . Van- peene, J. Villanova, S. Lyonnard, T. Vegge, and A. Bhowmik, “Utilizing active learning to accel- erate segmentation of microstructures with tiny annotation budgets,” Energy Storage Materials , vol. 73, p. 103785, 2024

  44. [54]

    Array programming with NumPy,

    C. R. Harris, K. J. Millman, S. J. van der Walt, R. Gommers, P. Virtanen, D. Courna- peau, E. Wieser, J. Taylor, S. Berg, N. J. Smith, R. Kern, M. Picus, S. Hoyer, M. H. van Kerkwijk, M. Brett, A. Haldane, J. F. del R´ıo, M. Wiebe, P. Peterson, P. G´erard-Marchant, K. Sheppard...

  45. [55]

    scikit-image: image processing in python,

    S. van der Walt, J. L. Sch ¨onberger, J. Nunez- Iglesias, F. Boulogne, J. D. Warner, N. Yager, E. Gouillart, and T. Yu, “scikit-image: image processing in python,” PeerJ, vol. 2, p. e453, June 2014

  46. [56]

    Heterogeneity of the Dominant Causes of Performance Loss in End-of-Life Cathodes and Their Consequences for Direct Recycling,

    M. Popeil, F. L. Usseglio-Viretta, X. Pu, P. Gasper, N. Dutta, E. Wang, E. Allen, J. S. Mangum, N. Sunderlin, K. Fink, J. M. Allen, P. J. Weddle, S. C. DeCaluwe, and D. P. Fine- gan, “Heterogeneity of the Dominant Causes of Performance Loss in End-of-Life Cathodes and Their Co...

  47. [57]

    Benchmarking feature upsampling methods for vision foundation models using interactive seg- mentation,

    V . Havrylov, H. Huang, D. Zhang, and A. Geiger, “Benchmarking feature upsampling methods for vision foundation models using interactive seg- mentation,” arXiv preprint, arXiv:2505.02075 , 2025

  48. [58]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haziza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal, H...

  49. [59]

    Doitpoms micrograph library,

    J. Eliot, “Doitpoms micrograph library,” 2000. https://www.doitpoms.ac.uk/index.php

  50. [60]

    Micrograph 394,

    R. F. Cochrane, “Micrograph 394,” 2002. https://www.doitpoms.ac.uk/miclib/ micrograph_record.php?id=394

  51. [61]

    Mosilib: Innovative anode materials for more powerful and sustainable batter- ies

    D. Cupid, “Mosilib: Innovative anode materials for more powerful and sustainable batter- ies.” https://www.ait.ac.at/en/themen/ battery-materials-characterization/ projects/mosilib

  52. [62]

    Denoising vision transformers,

    J. Yang, K. Z. Luo, J. Li, C. Deng, L. Guibas, D. Krishnan, K. Q. Weinberger, Y . Tian, and Y . Wang, “Denoising vision transformers,”arXiv preprint, arXiv:2401.02957, 2024. 14 Docherty et al. Feature Upsampling & Micrograph Segmentation Preprint Input channels, din 32/64/128/...

  53. [65]

    Gaussian blurs at each pixel for a set of strengthsσ∈{ 0, 1, 2, 4, 8, 16}

  54. [66]

    Sobel edge detection at each pixel for each of the Gaussian filtered arrays in 1

  55. [67]

    Hessian texture filter at each pixel for each of the Gaussian filtered arrays in 1, extracting the first two eigenvalues, as well as the mod, trace and determinant of the Hessian matrix

  56. [68]

    Difference of Gaussians of each of the Gaussian arrays in 1

  57. [69]

    S2 Hyperparameters The training hyperparameters for our upsampler is detailed in Table S1

    Membrane projections: convolution of the image array with a stack of line kernels oriented at 30 ◦ angle increments in [0◦, 180◦). S2 Hyperparameters The training hyperparameters for our upsampler is detailed in Table S1. The network architecture is shown in Figure S1. S3 Feat...

  58. [2017]

    Type: Journal Article

  59. [2024]

    Publisher: The Open Journal

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.