Pith. sign in

REVIEW 4 major objections 5 minor 20 references

LoRCA: LoRA Cycle Adaptation for Histology to HiP-CT Translation with DINOv3

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A shared frozen vision backbone with per-modality LoRA adapters produces structurally faithful histology-to-HiP-CT translation from unpaired data.

desk verdict Useful domain application of LoRA+CycleGAN, but the headline comparison against CycleGAN is not actually measured on translation quality. read the letter →

arxiv 2608.10002 v1 pith:ERYZ7NYD submitted 2026-08-07 eess.IV cs.CV

Yang Zhou , Edoardo Occhipinti , Banboye Kidzeru Elvis , Jishizhan Chen , Stathis Megas , Joseph Brunet , Joanna Purzycka , Theresa Urban
show 6 more authors
Hector Dejea Sarah Amalia Teichmann Menna R Clatworthy Paul Tafforeau Peter D Lee Claire L Walsh
This is my paper · ORCID
classification eess.IVcs.CV
keywords histology-to-HiP-CTtranslationcycle-consistentGANLoRAadaptersDINOv3Hicross-modalityregistrationunpairedimagestructuralconsistency
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

This paper tries to establish that a cycle-consistent translation framework built on a frozen DINOv3 vision backbone, with small modality-specific LoRA adapters, can convert H&E histology slides into images that look like HiP-CT scans without losing anatomical structure. The practical goal is to make 2D histology sections registerable to 3D HiP-CT volumes of intact human organs, a step that could give cellular-level staining context to whole-organ X-ray imaging. On its own test data, LoRCA improves Canny edge overlap over CycleGAN from about 0.19 to 0.42 and raises the number of reliable cross-modality feature matches from 62 to 1,761, a 28-fold increase. The authors see this as preliminary evidence that synthesis-before-registration for histology-to-HiP-CT is feasible without paired training data or manual anatomical contours.

What carries the argument

The central object is the shared frozen DINOv3 backbone with dual LoRA adapters. LoRA (low-rank adaptation) modifies a frozen linear layer's output by adding a trainable low-rank term $\frac{\alpha}{r}BAx$, so only small matrices $A$ and $B$ are learned per modality while the backbone weights $W$ stay fixed. Two independent LoRA sets, one for HiP-CT and one for H&E, are activated depending on the input modality, producing modality-specific feature maps from the same semantic feature space. A skip-connection UNet-style decoder reconstructs image space from the ViT patch tokens, and a cycle-consistent adversarial loss enforces that translating to the other modality and back recovers the original image.

What would settle it

Feature matches on one manually aligned pair are encouraging, but the claim of registration utility would be settled by measuring target registration error on several kidneys after running full slice-to-volume registration; if LoRCA-translated images do not reduce alignment error relative to grayscale or CycleGAN-translated inputs, the structural-fidelity gains would not translate to the stated purpose. Equally, an ablation that trains LoRCA with the DINOv3 backbone unfrozen or replaced by a random encoder would falsify the claim that the frozen anchor is the cause of the structural preservation.

Watch

Extended reading notes

Core claim

LoRCA is the first DINOv3-based cycle-consistent style translation framework for H&E histology to HiP-CT. The generator shares one frozen DINOv3 ViT as a structural anchor, injects separate low-rank LoRA adapters for the histology and HiP-CT modalities into all linear layers, and decodes each modality through its own UNet-style decoder, with PatchGAN discriminators and a cycle-consistency loss providing unpaired training. The central result is that this design preserves fine boundaries such as vessel walls and glomerular capsules better than CycleGAN, as measured by mutual information and Canny edge IoU, and that the translated images yield far more geometrically consistent keypoint matches under MatchAnything. The paper claims the frozen backbone prevents content drift by preserving pretrained semantic-extraction capability while the LoRA adapters handle modality-specific appearance.

Load-bearing premise

The method rests on the assumption that DINOv3, a model pretrained on natural images, already encodes the anatomical structures visible in both H&E histology and HiP-CT so strongly that a frozen shared backbone can serve as a reliable anchor for translating between them, with only small LoRA adapters needing to be trained.

Editorial extensions

If this is right

  • If LoRCA's design holds, unpaired histology-to-HiP-CT translation no longer needs manual anatomical contours, removing a barrier for soft-tissue organs.
  • The reported Canny edge IoU improvement from 0.188 (CycleGAN ResNet-9) to 0.417 indicates that structural boundaries are transferred rather than hallucinated, a necessary condition for registration.
  • The 28-fold increase in retained keypoint matches (62 to 1,761) suggests LoRCA-translated histology is a much better input for feature-based slice-to-volume registration than raw grayscale histology.
  • Because the backbone is frozen and only small adapters are trained, the method is cheap to adapt to new tissue types or new imaging modalities without retraining a full generator.
  • The cycle-consistency and PatchGAN losses, inherited from CycleGAN-style training, let the method work on completely unpaired datasets.

Reading between the lines

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

  • Beyond the paper's experiments, the same frozen-backbone-plus-LoRA cycle design could plausibly be applied to other cross-modal pairs where one modality lacks paired data, such as MRI to histology or EM to light microscopy, provided the semantic features transfer.
  • The paper stops at feature correspondences; if a full registration pipeline is added, the 28-fold match increase may or may not convert into lower target registration error, so the registration benefit remains an inference until measured.
  • A controlled ablation that unfreezes the backbone or replaces DINOv3 with a randomly initialized encoder would directly test whether the frozen semantic anchor, not just the increased parameter efficiency, causes the structural fidelity gain.
  • The dependence on DINOv3's natural-image pretraining means performance on organs or stains far outside the pretraining distribution is an open question; evaluating on a second organ type would test generalizability.
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

4 major / 5 minor

Summary. The paper proposes LoRCA, a cycle-consistent style translation framework for unpaired H&E histology to HiP-CT translation. The method uses a shared frozen DINOv3 backbone with modality-specific LoRA adapters, UNet-style decoders, PatchGAN discriminators, and a cycle-consistency objective. Evaluations are performed on human kidney data, reporting FID for three LoRCA decoder variants, mutual information and Canny edge IoU against CycleGAN baselines, and MatchAnything feature matching on one manually aligned test slice. The authors claim that LoRCA outperforms CycleGAN in both translation quality and structural consistency.

Significance. The problem is timely and important: enabling histology-to-HiP-CT synthesis would support slice-to-volume registration for an emerging imaging modality. The design of a shared frozen backbone with modality-specific LoRA adapters is a sensible way to inject semantic constraints into unpaired translation, and the paper usefully releases code and reports ethical approvals. However, the current evaluation only partially supports the headline. Structural consistency is quantified with self-consistency metrics, while the translation-quality component lacks a head-to-head FID comparison against CycleGAN, and the feature-matching evidence rests on a single slice. These gaps are fixable within the scope of a revision, so the contribution is promising but not yet fully substantiated.

major comments (4)
  1. [Abstract; §3.1; Table 1] The claim that LoRCA outperforms CycleGAN in translation quality is not supported by the reported metrics. Section 3.1 reports FID only for the three LoRCA decoder variants (plain 607.40, fusion 79.84, skip 77.18), and Table 1 reports no FID or KID for the CycleGAN ResNet-9 and UNet256 baselines. Since translation quality is one of the two components of the headline claim, the revision should compute FID (and ideally KID) between real HiP-CT patches and the outputs of both CycleGAN baselines and report them alongside LoRCA; without these numbers the superiority claim on translation quality is unsubstantiated.
  2. [§3.2, Table 1] The two metrics in Table 1, mutual information and Canny edge IoU, are computed between the generated HiP-CT patch and the input H&E patch, not between the generated output and real HiP-CT. Such metrics measure how well the source structure is preserved and can be maximized by a near-grayscale copy of the input; they do not establish that the output resembles real HiP-CT. Because the cycle-consistency loss in Eq. (3) enforces a related self-reconstruction objective, these numbers are partly self-referential. The paper should add target-domain fidelity metrics (for example, FID/KID against real HiP-CT or feature-statistics comparisons using a pretrained medical-imaging model) for all methods to support the translation-quality claim.
  3. [§3.3] The feature-matching evidence for downstream registration utility is based on a single manually aligned test slice. On that slice CycleGAN yields 19 retained matches versus 62 for grayscale histology and 1,761 for LoRCA; these absolute numbers suggest that keypoint-based matching is sensitive to slice content, tissue type, and manual alignment. To make the registration-utility claim robust, the authors should report the same MatchAnything experiment over multiple test slices or regions, with per-slice counts, and ideally compare against a CycleGAN output on the same slices. As it stands, a single-slice result cannot carry the downstream claim.
  4. [§4] The limitation paragraph acknowledges that backbone freezing has not been isolated by a controlled ablation, but it does not acknowledge the missing head-to-head FID comparison against CycleGAN. Since the abstract and Section 2.2 present the frozen DINOv3 backbone as the mechanism that prevents content drift, the lack of any ablation means the architectural claim is untested. The revision should include at least an ablation of LoRCA with an unfrozen backbone, or a from-scratch encoder, to show that the structural anchor is responsible for the reported improvements.
minor comments (5)
  1. [§2.2, Eq. (1)] The notation introduces b as a bias term, but the original linear layer's bias is not shown in Wx; please clarify whether b is a learnable LoRA-specific bias and whether the frozen layer has a separate bias.
  2. [§3.1] The FID is computed between real HiP-CT training patches and generated patches translated from histology; please specify whether the FID evaluation uses training or held-out patches and report the number of patches used.
  3. [§3.3] The text says CycleGAN performed worse, but its retained-match percentage (18.4%) is slightly higher than grayscale histology (17.4%); the statement should say fewer absolute retained matches to avoid ambiguity.
  4. [Table 1] The CycleGAN training and test values for Canny edge IoU are identical to three decimal places (0.188±0.024 for both splits, and 0.166±0.030 for both splits); please confirm the evaluation splits and patch sampling, as identical statistics across splits are unusual.
  5. [§2.4, Eq. (3)] The two expectations in the cycle-consistency loss use the same notation for the distribution over x_hist and x_hipct; writing explicit distributions would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified: the LoRCA derivation is self-contained; evaluation gaps are evidential limitations, not circular reductions.

full rationale

I walked the claimed derivation chain. The method (Eqs. 2-3) is a standard unpaired cycle-consistent GAN with a frozen DINOv3 backbone and modality-specific LoRA adapters; the losses are external objectives, not restatements of the evaluation metrics. The structural fidelity metrics in Table 1 (mutual information and Canny edge IoU between generated HiP-CT and input H&E) are related in spirit to the cycle-consistency objective but are not mathematically equivalent to Eq. (3): cycle consistency supervises round-trip reconstruction, whereas the metrics score direct edge/intensity preservation, and the adversarial term pushes outputs toward the HiP-CT distribution, so high scores are not forced by construction. The FID numbers in Sec. 3.1 are used only for LoRCA decoder selection; the absence of a CycleGAN FID comparison is an evidential gap in the 'translation quality' claim, not a circular reduction. The feature-matching experiment (Sec. 3.3) is an external downstream test using MatchAnything; it is not equivalent to the training loss. Self-citations [16] and [19] supply data and prior segmentation context, not a uniqueness theorem or a load-bearing theoretical premise. Section 4 explicitly acknowledges the missing registration-accuracy evaluation and the lack of a backbone-freezing ablation; these are limitations, not admissions of input-output equivalence. I therefore find no step in which a prediction reduces by construction to a fitted parameter or to a self-citation chain.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

There are no newly introduced physical entities. The framework uses existing components and hyperparameters chosen by hand without sensitivity analysis.

free parameters (7)
  • LoRA rank r = 8
    Chosen by hand; controls adapter capacity.
  • LoRA scaling factor alpha = 16
    Chosen by hand; scales adapter contributions.
  • lambda_adv = 5
    Relative weight of adversarial loss in total objective.
  • lambda_cyc = 10
    Relative weight of cycle consistency loss.
  • learning rate = 0.0003
    For LoRA adapters and discriminators.
  • patch size = 256x256
    Patches generated at 256 squared pixels, with tissue content threshold.
  • tissue content threshold = 90%
    Patches with less than 90% tissue discarded, affecting training distribution.
assumptions (4)
  • domain assumption DINOv3 pretrained on natural images provides semantic features transferable to H&E and HiP-CT modalities.
    The entire method relies on the shared frozen backbone to provide structure. Invoked in Section 2.2.
  • domain assumption Unpaired cycle-consistent adversarial training can learn a structure-preserving translation between these modalities.
    Assumed by the use of CycleGAN-style losses in Section 2.4.
  • domain assumption The manually aligned test H&E slice and HiP-CT slice are accurately aligned.
    The feature matching evaluation in Section 3.3 depends on this alignment being correct.
  • domain assumption Mutual information and Canny edge IoU between input and output measure structural fidelity.
    Used to support the claim of structural consistency in Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LoRCA: LoRA Cycle Adaptation for Histology to HiP-CT Translation with DINOv3." pith.science (2026). https://pith.science/paper/ERYZ7NYD

@misc{pith2026260810002,
  author       = {Pith},
  title        = {Pith review of: LoRCA: LoRA Cycle Adaptation for Histology to HiP-CT Translation with DINOv3},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ERYZ7NYD}},
  note         = {Machine review of arXiv:2608.10002}
}
abstract

Hierarchical Phase-Contrast Tomography (HiP-CT) is a synchrotron based X-ray imaging technique that enables non-destructive, volumetric imaging of intact organs with multi-resolutions bridging 20 $\mu m$/voxel for whole organs to near-cellular resolution ($\sim$0.8 $\mu m$/voxel) in local regions. This offers the opportunity to bring volumetric whole-organ context to histology. However, nonlinear registration between H\&E histology and HiP-CT volumes is challenging due to the differences in feature representations of different colour spaces. Synthesis-before-registration methods have shown strong results in histology-to-MRI and histology-to-CT alignment. However, existing approaches either rely on manual anatomical contours or are trained from scratch without semantic constraints, limiting their generalisability to soft tissue organs and novel modalities. We propose LoRCA (LoRA Cycle Adaptation), a cycle consistent style translation framework built on a shared frozen DINOv3 with modality-specific LoRA adapters, learning modality-specific representations that are decoded and adversarially trained. LoRCA enables structure-preserving translation without requiring paired training data. The frozen backbone is intended to be a structural anchor that prevents content drift by preserving pretrained semantic-extraction capability. We evaluate translation quality using Fr\'echet Inception Distance (FID) and structural fidelity via mutual information and Canny edge preservation. LoRCA outperforms CycleGAN in both translation quality and structural consistency. As a preliminary indicator of downstream registration utility, we find that style-translated images yield increased feature correspondences under MatchAnything on manually aligned HiP-CT and histology test pairs, suggesting that LoRCA-style translation is a promising step towards 2D histological sections to 3D HiP-CT volumes registration.

Figures

Figures reproduced from arXiv: 2608.10002 by the authors.

Figure 1
Figure 1. Overview of the pipeline of style translation between H&E histological sections and HiP-CT. (A) Illustration of organ scanning using HiP-CT, followed by shipping back for histological sectioning on the high-resolution VOIs. (B) Preparations of train￾ing data and the statistics of the four training samples. (C) LoRCA structure. (4) Application for registration. Serial sectioning was performed on two kidneys, generati… view at source ↗
Figure 2
Figure 2. LoRCA generator structure selection. (A) Illustration of different generator structures. (B) Generative HiP-CT from different models on H&E histology validation patches: vasculature wall (first row) and glomeruli (second row). where the adversarial loss Ladv follows the least-squares GAN formulation [10]. λadv and λcyc are the relative weights of the adversarial and cycle consistency terms. The cycle consistency los… view at source ↗
Figure 3
Figure 3. Comparison of edge preservation for CycleGAN and LoRCA on representative patches containing a blood vessel (top row) and a cortex with glomeruli (bottom row). Input Canny edge maps and edge overlap visualisations from generative ones are shown. recover spatial detail without hierarchical features. The fusion decoder substan￾tially improves realism (FID: 79.84), while the skip-connection decoder achieves the best FID… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Feature points searching and matching on manually aligned HiP-CT slice and histological WSI using MatchAnything [5]. glomeruli, are style-translated by each method. The Canny edge overlay com￾pares structural boundaries between the generated HiP-CT and the input H&E pa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 15 canonical work pages

  1. [1]

    arXiv preprint arXiv:2312.02366 (2023)

    Baharoon, M., Qureshi, W., Ouyang, J., Xu, Y., Aljouie, A., Peng, W.: Eval- uating General Purpose Vision Foundation Models for Medical Image Analysis: An Experimental Study of DINOv2 on Radiology Benchmarks. arXiv preprint arXiv:2312.02366 (2023)

    arXiv 2023

  2. [2]

    Light: Science & Applications 12(1), 57 (2023) 10 Y

    Bai, B., Yang, X., Li, Y., Zhang, Y., Pillar, N., Ozcan, A.: Deep learning-enabled virtual histological staining of biological samples. Light: Science & Applications 12(1), 57 (2023) 10 Y. Zhou et al

    work page 2023

  3. [3]

    In: International workshop on simulation and synthesis in medical imaging

    Casamitjana, A., Mancini, M., Iglesias, J.E.: Synth-by-reg (sbr): Contrastive learn- ing for synthesis-based registration of paired images. In: International workshop on simulation and synthesis in medical imaging. pp. 44–54. Springer (2021)

    work page 2021

  4. [4]

    Scientific Reports15(1), 25972 (2025)

    Chen,J.,Ronchetti,M.,Stehl,V.,Nguyen,V.,Kallaa,M.A.,Gedara,M.T.,Lölkes, C., Moser, S., Seidl, M., Wieczorek, M.: 2d-3d deformable image registration of histology slide and micro-ct with disa-based initialization. Scientific Reports15(1), 25972 (2025)

    work page 2025

  5. [5]

    arXiv preprint arXiv:2501.07556 (2025)

    He, X., Yu, H., Peng, S., Tan, D., Shen, Z., Bao, H., Zhou, X.: MatchAnything: Universal Cross-Modality Image Matching with Large-Scale Pre-Training. arXiv preprint arXiv:2501.07556 (2025)

    arXiv 2025

  6. [6]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-Rank Adaptation of Large Language Models. In: Proceedings of the International Conference on Learning Representations (ICLR) (2022)

    work page 2022

  7. [7]

    Medical Image Analysis50, 127–144 (2018)

    Iglesias, J.E., Modat, M., Peter, L., Stevens, A., Annunziata, R., Vercauteren, T., Lein, E., Fischl, B., Ourselin, S., Initiative, A.D.N., et al.: Joint registration and synthesis using a probabilistic model for alignment of MRI and histological sections. Medical Image Analysis50, 127–144 (2018)

    work page 2018

  8. [8]

    In: International Conference on Medical Image Computing and Computer- Assisted Intervention

    Leroy, A., Cafaro, A., Gessain, G., Champagnac, A., Grégoire, V., Deutsch, E., Lepetit, V., Paragios, N.: Structuregnet: Structure-guided multimodal 2d-3d regis- tration. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 771–780. Springer (2023)

    work page 2023

Show all 20 references
  1. [9]

    Li, Y., Wu, Y., Lai, Y., Hu, M., Yang, X.: MedDINOv3: How to adapt vision foun- dation models for medical image segmentation? arXiv preprint arXiv:2509.02379 (2025)

    2025

  2. [10]

    In: Proceedings of the IEEE International Conference on Computer Vision (ICCV)

    Mao, X., Li, Q., Xie, H., Lau, R.Y., Wang, Z., Smolley, S.P.: Least Squares Gener- ative Adversarial Networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2813–2821 (2017)

    2017

  3. [11]

    arXiv preprint arXiv:2304.07193 (2023)

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

    2023 arXiv

  4. [12]

    In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention

    Scholz, D., Erdur, A.C., Ehm, V., Meyer-Baese, A., Peeken, J.C., Rueckert, D., Wiestler, B.: MM-DINOv2: Adapting Foundation Models for Multi-modal Medical Image Analysis. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervent...

    2025

  5. [13]

    arXiv preprint arXiv:2508.10104 (2025)

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

    2025 arXiv

  6. [14]

    Cell Reports Methods3(9), 100592 (2023)

    Tasnadi, E., Sliz-Nagy, A., Horvath, P.: Structure preserving adversarial generation of labeled training samples for single-cell segmentation. Cell Reports Methods3(9), 100592 (2023)

    2023

  7. [15]

    Vasiljević, J., Nisar, Z., Feuerhake, F., Wemmert, C., Lampert, T.: Cyclegan for virtual stain transfer: Is seeing really believing? Artificial Intelligence in Medicine 133, 102420 (2022) LoRCA 11

    2022

  8. [16]

    Walsh, C.L., Tafforeau, P., Wagner, W.L., Jafree, D.J., Bellier, A., Werlein, C., Kühnel, M.P., Boller, E., Walker-Samuel, S., Robertus, J.L., Long, D.A., Jacob, J., Marussi, S., Brown, E., Holroyd, N., Jonigk, D.D., Ackermann, M., Lee, P.D.: Imaging intact human organs with l...

    2021

  9. [17]

    Computerized Medical Imaging and Graphics108, 102249 (2023)

    Wang, J., Wu, Q.J., Pourpanah, F.: DC-cycleGAN: Bidirectional CT-to-MR syn- thesis from unpaired data. Computerized Medical Imaging and Graphics108, 102249 (2023)

    2023

  10. [18]

    Computerized Medical Imaging and Graphics117, 102431 (2024)

    Wang, R., Heimann, A.F., Tannast, M., Zheng, G.: CycleSGAN: A cycle-consistent and semantics-preserving generative adversarial network for unpaired MR-to-CT image synthesis. Computerized Medical Imaging and Graphics117, 102431 (2024)

    2024

  11. [19]

    PLOS Computational Biology22(2), e1013923 (2026)

    Zhou, Y., Aslani, S., Javanmardi, Y., Brunet, J., Stansby, D., Carroll, S., Bellier, A., Ackermann, M., Tafforeau, P., Lee, P.D., et al.: Multiscale segmentation using hierarchical phase-contrast tomography and deep learning. PLOS Computational Biology22(2), e1013923 (2026)

    2026

  12. [20]

    In: Proceedings of the IEEE Inter- national Conference on Computer Vision (ICCV)

    Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. In: Proceedings of the IEEE Inter- national Conference on Computer Vision (ICCV). pp. 2242–2251 (2017)

    2017

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.