Foreseeing the Invisible: Amodal Reconstruction of Leaf Fossil Images

Ailin Zhang; Liuxiang Yue; Yikun Duan; Ziyue Zhao

REVIEW 5 major objections 5 minor 29 references

Foreseeing the Invisible: Amodal Reconstruction of Leaf Fossil Images

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that the missing tissue of a fossil leaf can be reconstructed from a single photograph without any visible-mask input, by fully fine-tuning a DINOv3 backbone and adding auxiliary venation heads.

desk verdict New task and strong public-benchmark results, but the headline fossil numbers rest on a contradiction in the ROI protocol that must be resolved. read the letter →

arxiv 2608.04423 v1 pith:B7GD2Y5Y submitted 2026-08-05 cs.CV

Liuxiang Yue , Ailin Zhang , Ziyue Zhao , Yikun Duan This is my paper · ORCID

classification cs.CV

keywords amodalsegmentationvisiontransformersdensepredictionpaleobotanyshapecompletionon-deviceinferencesyntheticdataleaffossilreconstruction

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 claims that the missing tissue in leaf-fossil photographs—the lamina hidden by rock, broken off, or eroded—can be recovered by treating the task as amodal reconstruction: predicting the full leaf and its veins from the visible pixels rather than just labelling what is exposed. The authors build AmodalDINO, a multi-head model that takes a single RGB image and outputs four masks (visible leaf, complete amodal leaf, primary vein, fine veins), with no visible-mask input at runtime. Trained only on 10,000 synthetic images rendered from 160 cleared-leaf specimens composited onto randomized stones, it reaches 95.0% Dice / 90.5% IoU on the amodal leaf in the held-out validation split and transfers qualitatively to nine real fossil photographs. Stripped to two heads and RGB input, the same recipe reports 85.05 full mIoU / 66.65 occluded mIoU on KINS and 80.90 / 38.15 on COCOA-cls, the KINS occluded figure being the highest in the paper's comparison table. If the transfer claim holds, paleobotanists gain an offline tool that turns partial fossils into measurable complete outlines and estimates leaf area from a ruler already present in the photograph.

What carries the argument

The engine is a DINOv3 ViT-L/16 transformer (a self-supervised vision transformer pretrained on the LVD-1689M weights) that the paper fully fine-tunes at a small learning rate instead of freezing, feeding a DPT decoder that fuses four intermediate token blocks into a dense feature map. Four independent $3\times3$ convolutional heads read out the visible leaf, the amodal leaf, the amodal primary vein, and the fine detail vein. The identity carrying the argument is that the two choices act on the same axis: unfreezing the backbone lets margins snap to the specimen's geometry, while supervising the venation heads forces the shared trunk to represent lobe and sinus structure, because a coherent midrib can only be drawn on a coherent lamina. An optional ROI channel, encoded with 0.5 as neutral so that an absent hint contributes exactly zero to the patch embedding, is an interface feature that plays no role in the reported numbers.

What would settle it

Have experts draw amodal leaf and vein ground truth on, say, fifty real fossil photographs, run the released model, and compare amodal-leaf Dice and IoU against the 95.0% Dice / 90.5% IoU reported on synthetic validation; a large drop would indict the synthetic-to-real premise rather than the architecture. As a second arm, retrain the frozen-backbone ablation on the same real labels: if it matches the unfrozen model there, the claimed benefit of full fine-tuning does not survive contact with real data.

Watch

Extended reading notes

Core claim

AmodalDINO's central claim is that the missing tissue of a fossil leaf is recoverable from the visible pixels alone, and that two training decisions make the recovery specimen-specific rather than generic. The paper shows that a DINOv3 ViT-L/16 backbone fully fine-tuned at a learning rate of $10^{-5}$ rather than frozen, with a DPT decoder and four independent heads (visible leaf, amodal leaf, amodal primary vein, fine detail vein), learns the structural shape prior of leaves from purely synthetic data. On the held-out synthetic validation split the amodal leaf head reaches 95.0% Dice and 90.5% IoU, an 8.0 IoU gap below the near-saturated visible head that the authors read as the portion of the task that is genuine hallucination. Without any visible-mask input, the two-head variant reports 85.05 full mIoU / 66.65 occluded mIoU on KINS and 80.90 / 38.15 on COCOA-cls, with the KINS occluded number the highest in the comparison table. The same model, quantized to 4-bit weights, agrees with the full-precision checkpoint at 0.910 IoU on the amodal leaf.

Load-bearing premise

The premise that carries the whole result is that the synthetic training set—rendered in Blender from 160 hand-picked cleared-leaf specimens composited onto randomized stones—spans the visual variability of real fossil leaves well enough that a model trained only on synthetic images completes real specimens, and the paper's direct evidence for that transfer is qualitative (nine photographs).

Editorial extensions

If this is right

Amodal reconstruction can be done without any visible-mask input, so a deployment pipeline needs no upstream instance segmenter at runtime.
Fully fine-tuning a self-supervised vision transformer on a small synthetic dataset preserves generalization while specializing the completion to specimen-specific shape.
Auxiliary structural supervision—predicting venation alongside the leaf—acts as an inductive bias that direct supervision of the target mask alone does not provide.
The same recipe, stripped to two heads, transfers to everyday amodal segmentation: it reaches the highest occluded mIoU in the paper's KINS comparison without seeing a visible mask.
Quantizing the model to 4-bit weights keeps amodal-leaf agreement at 0.910 IoU, allowing the full pipeline to run offline in a browser.

Reading between the lines

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

Beyond the paper: because the model completes organic shapes without a category template, the same fine-tuning-plus-structural-heads recipe is a plausible starting point for other non-rigid objects with internal structure, such as broken bones, roots, or archaeological fragments.
Beyond the paper: the KINS result suggests the field's standard protocol—feeding a visible mask to grow outward—may understate what a fully fine-tuned backbone can extrapolate; a controlled comparison with identical crops, backbones, and schedule would isolate that effect.
Beyond the paper: the synthetic-to-real transfer claim currently rests on nine qualitative photographs; a quantitative field test with expert-drawn amodal masks on a few dozen real specimens would turn the claim into a measured one.
Beyond the paper: the two-rate schedule ($10^{-5}$ backbone, $10^{-4}$ trunk) may or may not generalize to other self-supervised backbones, so testing whether the staging is DINOv3-specific would clarify whether the finding is about leaves or about transformer fine-tuning.

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.

Desk Editor's Note

read the letter

Two things to know about arXiv:2608.04423. First, the task is genuinely new—amodal reconstruction of leaf fossils from RGB alone, with the visible mask treated as an output rather than an input—and the KINS/COCOA results, obtained with a two-head RGB-only variant, are interesting in their own right. Second, the headline fossil numbers (95.0 Dice / 90.5 IoU) are tainted by an internal contradiction in Appendix B: the paper claims every fossil number is measured with a neutral hint, but then says validation uses the stone mask where available. Since every synthetic validation image has a stone mask, Table I likely was produced with a stone-mask ROI input. That is a substantial extra cue—it tells the model which pixels are background—and it undercuts the "no visible-mask" claim. The paper does not report a neutral vs stone-hinted comparison, so we cannot tell how much inflation this causes. This is not a judgment about the synthetic-to-real gap; it is a protocol inconsistency that must be fixed before the fossil numbers can be interpreted.

What the paper does well: the task framing is useful, the synthetic data pipeline is thoughtful (160 hand-picked leaves, Blender compositing, breakage variants), and the ablations on full fine-tuning and auxiliary vein heads are honestly reported and show real effects. The deployment details—4-bit quantization, browser demo, ruler calibration—are practical and reproducible in spirit, though no code or data are released.

Soft spots: the checkpoint-selection issue the reader flagged is real but mild; reporting the maximum over validation-checkpoint epochs on the same split inflates Dice by a few points. Real-fossil generalization rests on qualitative inspection of nine photographs, which the paper admits. The KINS/COCOA numbers are more trustworthy because the ROI channel is dropped entirely; still, the crop-centering convention does part of the instance-identification work, and the protocol differs from prior methods, so calling the KINS occluded number "highest" is fair only with that caveat.

This paper deserves a serious referee. The task and the no-mask recipe are interesting enough that a careful review could turn it into a solid contribution. The referee should insist on a resolution of the ROI protocol: re-report the fossil validation with neutral hint, or publish both numbers. Without that, the central claim does not stand. With it, this becomes a useful paper for paleobotany and amodal segmentation.

Referee Report

5 major / 5 minor

Summary. The paper proposes AmodalDINO, a multi-head dense-prediction model for amodal reconstruction of leaf fossils from a single RGB image. It is trained entirely on a synthetic dataset generated from 160 cleared-leaf specimens composited in Blender and augmented with breakage masks, and it predicts visible leaf, amodal leaf, amodal primary vein, and fine vein masks without being given a visible mask as input. The main reported result is 95.0% Dice / 90.5% IoU for the amodal leaf on the synthetic validation split (Table I), plus a two-head RGB-only variant that reaches 85.05 full mIoU / 66.65 occluded mIoU on KINS test and 80.90 / 38.15 on COCOA-cls val (Table II), and a 4-bit quantized browser deployment with ruler-based area calibration and an optional generative leaf visualization. The paper attributes the success to fully fine-tuning a DINOv3 ViT-L/16 backbone instead of freezing it, and to auxiliary venation heads that provide a structural shape prior.

Significance. The application is novel and potentially useful: no previous work addresses amodal completion of fossil leaves, and the no-visible-mask formulation is a meaningful departure from the standard amodal-segmentation protocol, even though the benchmark variant still uses ground-truth-box crops as a localization cue. If the results are reproducible, the synthetic-to-real transfer claim would be of interest to paleobotanists and to the amodal segmentation community. The paper is highly transparent: Appendix A lists every hyperparameter and loss coefficient, Appendix C describes the data pipeline and splits, the KINS test-set evaluation is independent, and the limitations section candidly lists the one-leaf assumption, the synthetic-to-real domain gap, and the qualitative nature of real-fossil evaluation. The 4-bit quantization fidelity check (Table III) is a concrete practical contribution. The main caveats are that the headline fossil validation number appears to be computed with a stone-mask ROI hint rather than the neutral hint, that checkpoint selection is performed on the same split that is then reported, and that the real-fossil transfer evidence is qualitative only.

major comments (5)

[Appendix B; Section III-A; Table I] The last paragraph of Appendix B states 'At validation the hint is the stone mask where available and neutral otherwise,' which contradicts Section III-A and the first paragraph of Appendix B, both of which say that every fossil number in Section V-C is measured with a neutral hint. Because the synthetic validation split always has a stone mask (Appendix C lists stone_mask under val), Table I's 95.0% Dice / 90.5% IoU are apparently computed with the stone mask supplied as the fourth input channel, not with the neutral hint. The stone mask is a non-RGB mask cue: it identifies the rock region, which is both the main false-positive source and the region where the occluded tissue lies, so it can materially inflate amodal performance. Please report validation metrics under both the neutral hint and the stone-mask hint, and state explicitly which protocol produced each number in Table I.
[Section V-A; Table I] The reported 95.0% Dice / 90.5% IoU is the best validation amodal Dice over checkpoints: Section V-A says 'We select the checkpoint by best validation amodal Dice, at epoch 37,' and Table I is evaluated on the same validation split. Selecting the checkpoint on the split that is then reported as the final result makes the headline number the maximum over 40 epochs and is optimistically biased. Please report performance on a truly held-out split (or use cross-validation) together with the selected-checkpoint number, or at least report the epoch-40 and validation-selected numbers side by side.
[Appendix C; Table II] For the public-benchmark variant, Appendix C states that the best checkpoint is chosen by validation full mIoU on each dataset ('epoch 26 on KINS, epoch 22 on COCOA-cls'), while Table II reports the COCOA-cls numbers on the COCOA-cls validation set. Thus the 80.90 / 38.15 COCOA-cls figures are selected and evaluated on the same split. The KINS numbers are cleaner because the KINS row reports the full test set with selection on a 10k-instance subset; please clarify the COCOA-cls protocol and, if possible, evaluate on a test split or quantify the selection bias.
[Section V-D; Figure 5] The claim that full fine-tuning and the four-head setup are the decisive design choices is supported only by qualitative examples in Figure 5. Since the validation split exists, please add a table with Dice/IoU for the full model, the frozen-backbone variant, and the single-head variant under identical training conditions; this would both strengthen the central method claim and make the comparison with Table I interpretable.
[Section V-F; Abstract] The statement that the model 'transfers well to real fossil specimens' rests on nine photographs with no ground truth, evaluated qualitatively. This is a severe evidentiary limitation for a generalization claim made in the abstract. Either add a quantitative evaluation on real specimens (even an expert-rated or proxy metric), or temper the claim to something like 'qualitative transfer on nine photographs.'

minor comments (5)

[Abstract] The abstract contains a typo: 'thecompleteshape' should be 'the complete shape'.
[Section III-C] The sentence after Eq. (1) reads 'Note thatwhere is a per-pixelweight map'; it should read 'Note that w is a per-pixel weight map'.
[Section V-D; Figure 5] Figure 5 would be easier to read if the rows were labeled with the exact configuration (frozen vs. unfrozen, which heads) in the figure itself; currently the caption describes rows and the reader must map back.
[Section VII] The limitation section says 'Further improvements includes query- or prompt-based heads'; the grammar should be 'include'.
[Section II; Section VI-A; Reference [21]] The text refers to 'YOLO26-seg' in several places, but the reference [21] is titled 'YOLOE: Real-time seeing anything'; please reconcile the naming.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the Appendix B ROI-protocol contradiction is a non-circular reporting flaw.

full rationale

Walked the claimed derivation chain. The synthetic-fossil result rests on a train/val split by scene (Section V-C, Appendix C), so no leaf instance or stone appears on both sides; the reported Dice/IoU is a measured overlap on held-out renders, not a quantity defined in terms of the training targets. The public-benchmark results are the most independent: Section V-E states the stripped variant is 'RGB-only, no ROI channel', trained separately on KINS and COCOA-cls, and evaluated on the KINS test set and the COCOA-cls validation set; the occluded-mIoU metric (Eq. 5) uses the ground-truth visible mask only as an evaluation device. No fitted parameter is later renamed as a prediction, and there are no load-bearing self-citations: the reference list contains no prior work by these authors, so the 'uniqueness' or 'adopted ansatz' patterns do not arise. Two caveats are real but are not circularity. First, Section V-A selects the checkpoint by 'best validation amodal Dice, at epoch37', so the headline 95.0 Dice is an optimistically selected validation statistic rather than an unbiased estimate; that is model selection, not a reduction by construction. Second, Appendix B's last sentence, 'At validation the hint is the stone mask where available and neutral otherwise', contradicts the repeated claim that 'every fossil number in the paper is measured with a neutral hint'; because every synthetic render emits a stone mask (Section IV), the fossil validation split may have been evaluated with a non-neutral, stone-mask ROI. This is a serious internal inconsistency that undercuts the 'neutral hint' description of the 95.0/90.5 numbers, but it does not make the derivation circular: the stone mask is not the target, and the same no-mask capability is independently demonstrated by the RGB-only KINS/COCOA runs. The paper is therefore self-contained against external benchmarks; the contradiction should be resolved as a correctness issue, not a circularity.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical machine learning study. The main free parameters are hyperparameters and the checkpoint selection epoch, which is fitted to the validation set. The axioms are domain assumptions about synthetic-to-real transfer and the use of standard benchmarks as proxies. No new physical entities are introduced.

free parameters (6)

Backbone learning rate = 1e-5
Section III-B and V-A: chosen because 1e-4 causes instability; tuned specifically for DINOv3.
Trunk/head learning rate = 1e-4
Section V-A: two-rate schedule is described as the narrowest part of the recipe.
Loss weight vector alpha = (1.0, 1.0, 1.0, 0.4)
Appendix A, Table IV: head weights chosen by hand, detail vein down-weighted due to low label quality.
Tversky beta parameters = beta_FP=0.65, beta_FN=0.35
Section III-C: tuned to favor precision over recall.
Checkpoint selection epoch = epoch 37
Section V-A: best validation amodal Dice, and the same validation split is then reported as the headline result.
Completion reweight lambda_c = 2.0
Eq 3, chosen to up-weight the occluded region.

assumptions (3)

domain assumption The synthetic dataset built from 160 cleared-leaf specimens captures real fossil variability.
Section IV: training is fully synthetic; real-fossil test is 9 photographs with no quantitative metrics, so transfer is assumed.
domain assumption Fully fine-tuning DINOv3 at a low learning rate preserves pretrained generalization.
Section V-D: the paper asserts small step size adapts features rather than overwriting them; no out-of-distribution benchmark is run.
domain assumption Standard amodal benchmarks (KINS, COCOA-cls) are appropriate proxies for the fossil task despite category differences.
Section V-E: used to validate the backbone/decoder recipe, but the benchmark instances are rigid common objects, not organic leaves.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Foreseeing the Invisible: Amodal Reconstruction of Leaf Fossil Images." pith.science (2026). https://pith.science/paper/B7GD2Y5Y

@misc{pith2026260804423,
  author       = {Pith},
  title        = {Pith review of: Foreseeing the Invisible: Amodal Reconstruction of Leaf Fossil Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B7GD2Y5Y}},
  note         = {Machine review of arXiv:2608.04423}
}

read the original abstract

Fossil leaves are rarely preserved whole -- sedimentary rock hides, breaks, and erodes the lamina, yet paleobotany depends on the complete shape and outline of the leaf. We cast the recovery of the missing tissue as amodal reconstruction and present AmodalDINO, a multi-head dense-prediction model that predicts four masks from a single RGB image: visible leaf, amodal complete leaf, amodal main vein, and fine veins. Unlike essentially all prior amodal work, AmodalDINO is given no visible mask. It predicts the visible and amodal regions jointly, so it needs no upstream instance segmenter at runtime. Two simple but effective changes adapt the model to the amodal segmentation task: fully fine-tune a DINOv3 ViT-L/16 at a small learning rate instead of freezing it, and attach auxiliary venation heads alongside the leaf heads. These two changes enable the model to learn the structural shape prior of leaves. Trained only on synthetic leaf fossil images, AmodalDINO reaches 95.0% Dice / 90.5% IoU on the validation set and transfers well to real fossil specimens. Stripped to two heads, the same recipe can run on two benchmark datasets, reaching 85.05 full mIoU / 66.65 occluded mIoU on KINS and 80.90 / 38.15 on COCOA-cls. The model is also practical: by quantizing to 4-bit weights, it runs entirely offline in a browser, matching the original model with an IoU of 0.910. We also add ruler-based calibration to estimate surface area, and a generative visualization of living leaves on local devices.

Figures

Figures reproduced from arXiv: 2608.04423 by the authors.

**Figure 2.** Top: the deployed system. A photograph of a specimen is segmented into stone and ruler, the stone cut-out is passed to AmodalDINO, and the four [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Synthetic leaf fossil images used for training. Pose, scale, lighting, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Soft outputs of the four heads for one specimen (left to right: input [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 7.** Figure 7: Generalization to real fossil specimens (photograph, predicted masks, [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 6.** Figure 6: The two-head, RGB-only variant on occluded KINS instances. From [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 8.** Figure 8: The offline browser demo, running locally in WebGPU on the 4-bit [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Failure mode: two leaf fossils on one slab. The amodal head (pale [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

**Figure 10.** Figure 10: The Blender compositing scene. A segmented cleared-leaf texture [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 21 canonical work pages

[1]

Computer vision cracks the leaf code,

P. Wilf, S. Zhang, S. Chikkerur, S. A. Little, S. L. Wing, and T. Serre, “Computer vision cracks the leaf code,”Proceedings of the National Academy of Sciences (PNAS), vol. 113, no. 12, pp. 3305–3310, 2016

work page 2016
[2]

Amodal instance segmen- tation with KINS dataset,

L. Qi, L. Jiang, S. Liu, X. Shen, and J. Jia, “Amodal instance segmen- tation with KINS dataset,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 3014– 3023

work page 2019
[3]

Semantic amodal segmen- tation,

Y . Zhu, Y . Tian, D. Metaxas, and P. Doll ´ar, “Semantic amodal segmen- tation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1464–1472

work page 2017
[4]

Sim ´eoni, H

O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquabet al., “DINOv3,”arXiv preprint arXiv:2508.10104, 2025

arXiv 2025
[5]

Self-supervised scene de-occlusion,

X. Zhan, X. Pan, B. Dai, Z. Liu, D. Lin, and C. C. Loy, “Self-supervised scene de-occlusion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 3784– 3792

work page 2020
[6]

Amodal segmentation based on visible region segmentation and shape prior,

Y . Xiao, Y . Xu, Z. Zhong, W. Luo, J. Li, and S. Gao, “Amodal segmentation based on visible region segmentation and shape prior,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 35, no. 4, 2021, pp. 2995–3003

work page 2021
[7]

Coarse- to-fine amodal segmentation with shape prior,

J. Gao, X. Qian, Y . Wang, T. Xiao, T. He, Z. Zhang, and Y . Fu, “Coarse- to-fine amodal segmentation with shape prior,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 1262–1271

work page 2023
[8]

AISFormer: Amodal instance segmentation with transformer,

M. Tran, K. V o, K. Yamazaki, A. Fernandes, M. Kidd, and N. Le, “AISFormer: Amodal instance segmentation with transformer,” inBritish Machine Vision Conference (BMVC), 2022

work page 2022

Show all 29 references

[9]

Learnable shape prototypes with occlusion-geometry-guided injection for amodal instance segmentation,

F. Zhang, J. Wang, and X. Ye, “Learnable shape prototypes with occlusion-geometry-guided injection for amodal instance segmentation,” arXiv preprint arXiv:2605.24533, 2026

2026 arXiv
[10]

PLUG: Revisiting amodal segmentation with foundation model and hierarchical focus,

Z. Liu, L. Qiao, X. Chu, and T. Jiang, “PLUG: Revisiting amodal segmentation with foundation model and hierarchical focus,”arXiv preprint arXiv:2405.16094, 2024

2024 arXiv
[11]

Segment anything,

A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Girshick, “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 4015–4026

2023
[12]

Amodal SAM: A unified amodal segmentation framework with generalization,

B. Zhang, Z. Tian, X. Tao, S. Tang, J. Yu, and W. Pei, “Amodal SAM: A unified amodal segmentation framework with generalization,”arXiv preprint arXiv:2604.20748, 2026

2026 arXiv
[13]

pix2gestalt: Amodal segmentation by synthesizing wholes,

E. Ozguroglu, R. Liu, D. Sur ´ıs, D. Chen, A. Dave, P. Tokmakov, and C. V ondrick, “pix2gestalt: Amodal segmentation by synthesizing wholes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 3931–3940

2024
[14]

Open-world amodal appearance completion,

J. Ao, Y . Jiang, Q. Ke, and K. A. Ehinger, “Open-world amodal appearance completion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 6490– 6499

2025
[15]

Amodal ground truth and completion in the wild,

G. Zhan, C. Zheng, W. Xie, and A. Zisserman, “Amodal ground truth and completion in the wild,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 28 003– 28 013

2024
[16]

DINOv2: Learning robust visual features without supervision,

M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov et al., “DINOv2: Learning robust visual features without supervision,” Transactions on Machine Learning Research (TMLR), 2024

2024
[17]

Vision transformers for dense prediction,

R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 12 179–12 188

2021
[18]

SAIN: Structure- aware image inpainting for large missing areas,

D. Wang, Y . Kang, Y . Chen, Y . Gao, and S. Xu, “SAIN: Structure- aware image inpainting for large missing areas,”Journal of King Saud University – Computer and Information Sciences, vol. 38, no. 4, 2026

2026
[19]

GPTQ: Accurate post-training quantization for generative pre-trained transformers,

E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh, “GPTQ: Accurate post-training quantization for generative pre-trained transformers,” in International Conference on Learning Representations (ICLR), 2023

2023
[20]

ONNX Runtime: Cross-platform, high- performance ml inferencing and training accelerator,

ONNX Runtime developers, “ONNX Runtime: Cross-platform, high- performance ml inferencing and training accelerator,” https://onnxrunt ime.ai, 2021

2021
[21]

YOLOE: Real-time seeing anything,

A. Wang, L. Liu, H. Chen, Z. Lin, J. Han, and G. Ding, “YOLOE: Real-time seeing anything,”arXiv preprint arXiv:2503.07465, 2025

2025
[22]

MobileCLIP: Fast image-text models through multi-modal reinforced training,

P. K. A. Vasu, H. Pouransari, F. Faghri, R. Vemulapalli, and O. Tuzel, “MobileCLIP: Fast image-text models through multi-modal reinforced training,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 15 963–15 974

2024
[23]

V-Net: Fully convolutional neural networks for volumetric medical image segmentation,

F. Milletari, N. Navab, and S.-A. Ahmadi, “V-Net: Fully convolutional neural networks for volumetric medical image segmentation,” inInter- national Conference on 3D Vision (3DV), 2016, pp. 565–571

2016
[24]

Tversky loss function for image segmentation using 3D fully convolutional deep networks,

S. S. M. Salehi, D. Erdogmus, and A. Gholipour, “Tversky loss function for image segmentation using 3D fully convolutional deep networks,” inInternational Workshop on Machine Learning in Medical Imaging (MLMI), 2017, pp. 379–387

2017
[25]

SAM 3: Segment anything with concepts,

N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, P. Dollar, N. Ravi, K. Saenko, P. Zhang, and C. Feichtenhofer, “SAM 3: Segment anything with concepts,”arXiv preprint arXiv:2511.16719, 2025

2025 arXiv
[26]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inInternational Conference on Learning Representations (ICLR), 2019

2019
[27]

PVT v2: Improved baselines with pyramid vision transformer,

W. Wang, E. Xie, X. Li, D.-P. Fan, K. Song, D. Liang, T. Lu, P. Luo, and L. Shao, “PVT v2: Improved baselines with pyramid vision transformer,” Computational Visual Media, vol. 8, no. 3, pp. 415–424, 2022

2022
[28]

EV A-02: A visual representation for neon genesis,

Y . Fang, Q. Sun, X. Wang, T. Huang, X. Wang, and Y . Cao, “EV A-02: A visual representation for neon genesis,”Image and Vision Computing, vol. 149, p. 105171, 2024

2024
[29]

FLUX.2: Frontier visual intelligence,

Black Forest Labs, “FLUX.2: Frontier visual intelligence,” https://bfl.ai /blog/flux-2, 2025. APPENDIXA FULLIMPLEMENTATIONDETAILS Table IV lists every hyperparameter used to train the released checkpointamodal_dino_vitl16_roi_ full_indep.pt. The reference implementation lives ...

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Computer vision cracks the leaf code,

P. Wilf, S. Zhang, S. Chikkerur, S. A. Little, S. L. Wing, and T. Serre, “Computer vision cracks the leaf code,”Proceedings of the National Academy of Sciences (PNAS), vol. 113, no. 12, pp. 3305–3310, 2016

work page 2016

[2] [2]

Amodal instance segmen- tation with KINS dataset,

L. Qi, L. Jiang, S. Liu, X. Shen, and J. Jia, “Amodal instance segmen- tation with KINS dataset,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 3014– 3023

work page 2019

[3] [3]

Semantic amodal segmen- tation,

Y . Zhu, Y . Tian, D. Metaxas, and P. Doll ´ar, “Semantic amodal segmen- tation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1464–1472

work page 2017

[4] [4]

Sim ´eoni, H

O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquabet al., “DINOv3,”arXiv preprint arXiv:2508.10104, 2025

arXiv 2025

[5] [5]

Self-supervised scene de-occlusion,

X. Zhan, X. Pan, B. Dai, Z. Liu, D. Lin, and C. C. Loy, “Self-supervised scene de-occlusion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 3784– 3792

work page 2020

[6] [6]

Amodal segmentation based on visible region segmentation and shape prior,

Y . Xiao, Y . Xu, Z. Zhong, W. Luo, J. Li, and S. Gao, “Amodal segmentation based on visible region segmentation and shape prior,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 35, no. 4, 2021, pp. 2995–3003

work page 2021

[7] [7]

Coarse- to-fine amodal segmentation with shape prior,

J. Gao, X. Qian, Y . Wang, T. Xiao, T. He, Z. Zhang, and Y . Fu, “Coarse- to-fine amodal segmentation with shape prior,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 1262–1271

work page 2023

[8] [8]

AISFormer: Amodal instance segmentation with transformer,

M. Tran, K. V o, K. Yamazaki, A. Fernandes, M. Kidd, and N. Le, “AISFormer: Amodal instance segmentation with transformer,” inBritish Machine Vision Conference (BMVC), 2022

work page 2022