{"id":"cf8670c8-49cd-40df-b15d-17d8ac2e04a0","arxiv_id":"2608.04423","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A model with a fully fine-tuned DINOv3 backbone and auxiliary vein heads performs amodal leaf reconstruction without any visible-mask input, reaching 95.0% Dice on synthetic fossils and top occluded mIoU on KINS.","lead":"AmodalDINO reconstructs the complete shape and venation of fossil leaves from a single photograph, without being told which pixels are visible. It reports 95% Dice on synthetic validation and the highest occluded mIoU on KINS among compared methods, suggesting the approach may also help other amodal segmentation tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix B contradicts the neutral-ROI claim: validation may use a stone-mask hint, so Table I's 95.0 Dice / 90.5 IoU may not be a no-mask result.","rationale":"The reader's weakest assumption is the synthetic-to-real domain gap, which is a real limitation but secondary: the paper already says real-fossil transfer is qualitative only, so no quantitative generalization claim is made beyond synthetic validation. The more load-bearing issue is an internal contradiction about the ROI protocol. Section III-A and Appendix B's first paragraph state that every fossil number uses a neutral ROI, yet Appendix B's final paragraph says the validation hint is the stone mask where available. Since every synthetic validation image has a stone mask, the headline 95.0 Dice / 90.5 IoU may have been produced with an extra stone-mask input, not with pure RGB. This is not a matter of domain disagreement; it is a verification failure that directly affects the central quantitative result. The ablations in Section V-D show relative effects but do not establish the absolute numbers, and Table III only compares 4-bit against fp32 on the same model, so it does not validate the reconstruction itself. The public-benchmark section also depends on ground-truth visible-box crops, but that is disclosed in Section V-E; the ROI contradiction is not. A single reproduction of the validation protocol with neutral versus stone-mask ROI would settle whether the headline numbers stand as claimed. I am not alleging deliberate misreporting; the discrepancy may be a documentation error, and the proposed test would distinguish between a harmless typo and an invalidated central claim.","tokens_in":15830,"tokens_out":6510,"duration_ms":68350,"concrete_test":"Load the released checkpoint amodal_dino_vitl16_roi_full_indep.pt and run the 1,000-image synthetic validation split twice: once with the ROI channel set to the neutral value 0.5 (normalized to zero) and once with the stone mask as the ROI, keeping all other settings identical. Compare per-head Dice/IoU against Table I. If the neutral-ROI run differs materially (say >1 IoU point on the amodal leaf) the paper must re-report Table I as stone-hinted and qualify the no-mask claim; if the runs match, the Appendix B validation sentence should be corrected to say neutral and the code's validation hint source checked to confirm it is not reading stone_mask.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central synthetic-fossil numbers are supposed to be measured with no ROI input: Section III-A says 'all fossil numbers in Section V-C are measured that way' (neutral hint), and Appendix B's first paragraph repeats 'every fossil number in the paper is measured with a neutral hint.' But the last paragraph of Appendix B states: 'At validation the hint is the stone mask where available and neutral otherwise.' The synthetic validation split always has a stone mask, so Table I's 95.0% Dice / 90.5% IoU may have been computed with the stone mask as a fourth input channel, not with neutral ROI. This matters because a stone mask reveals exactly which pixels are background/rock; the model can then restrict completion to the stone region, which is a substantial extra cue and is itself a non-RGB mask input. If actually used, it undercuts the 'no visible-mask input' headline and inflates the reported amodal numbers by an unknown amount. The paper reports no comparison between neutral and stone-hinted validation, so the reader cannot tell which protocol produced the headline numbers. This is an internal inconsistency, not a judgment about the synthetic-to-real gap; it must be resolved before the validation result can be interpreted.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":16092,"tokens_out":11782,"duration_ms":119574,"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":[{"comment":"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":"Appendix B; Section III-A; Table I"},{"comment":"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.","section":"Section V-A; Table I"},{"comment":"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":"Appendix C; Table II"},{"comment":"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":"Section V-D; Figure 5"},{"comment":"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.'","section":"Section V-F; Abstract"}],"minor_comments":[{"comment":"The abstract contains a typo: 'thecompleteshape' should be 'the complete shape'.","section":"Abstract"},{"comment":"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":"Section III-C"},{"comment":"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":"Section V-D; Figure 5"},{"comment":"The limitation section says 'Further improvements includes query- or prompt-based heads'; the grammar should be 'include'.","section":"Section VII"},{"comment":"The text refers to 'YOLO26-seg' in several places, but the reference [21] is titled 'YOLOE: Real-time seeing anything'; please reconcile the naming.","section":"Section II; Section VI-A; Reference [21]"}],"recommendation":"major_revision","confidential_remarks":"The most important issue is the internal inconsistency in Appendix B: if the authors confirm that validation uses the stone-mask hint, Table I must be recomputed and the headline claims revised. I recommend a revision that resolves the ROI protocol, reports unbiased validation numbers, and adds quantitative ablations. The paper is transparent about its limitations, which is a strength. Given the number of 2026 preprints cited and the paper's date, the editorial office may also want to verify the availability of those references."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat 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.\n\nSoft 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.\n\nThis 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.","headline":"New task and strong public-benchmark results, but the headline fossil numbers rest on a contradiction in the ROI protocol that must be resolved.","tokens_in":16601,"tokens_out":3134,"would_cite":false,"duration_ms":31885,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["amodal segmentation","vision transformers","dense prediction","paleobotany","shape completion","on-device inference","synthetic data","leaf fossil reconstruction"],"falsifier":"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.","tokens_in":15646,"feed_emoji":"🍃","tokens_out":8653,"duration_ms":77822,"temperature":0.7,"pith_summary":"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.","feed_headline":"No visible mask: model rebuilds whole leaf from fossil photo","feed_subtitle":"Trained only on synthetic images, it recovers buried leaf tissue and veins, and runs offline in a browser.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Provides the DINOv3 ViT-L/16 backbone with LVD-1689M weights that the method fully fine-tunes.","marker":"[4]"},{"why":"Supplies the DPT trunk that fuses four intermediate token blocks into the dense features read out by the four heads.","marker":"[17]"},{"why":"Defines the KINS dataset and its full/occluded mIoU protocol used for the public-benchmark comparison.","marker":"[2]"},{"why":"Defines the COCOA/COCOA-cls dataset and protocol used as the second public amodal benchmark.","marker":"[3]"},{"why":"SAM 3 is used to segment the 160 cleared-leaf seed specimens into complete lamina silhouettes.","marker":"[25]"},{"why":"Supplies the soft Dice term used in the visible and detail-vein head losses.","marker":"[23]"},{"why":"Supplies the Tversky loss used in the completion heads to penalize over-growth into stone more than missed pixels.","marker":"[24]"},{"why":"GRASP is the comparison baseline that, like this work, builds on a DINO backbone but keeps it frozen and consumes a visible mask.","marker":"[9]"}],"fun_headline_variants":["No mask needed: AI rebuilds fossil leaves from photo","Synthetic-only training rebuilds whole fossil leaf","Fossil leaf recovery: no visible mask, just RGB","AmodalDINO: rebuilding fossil leaves without a mask"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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).","fun_headline_variants_meta":{"raw":{"variants":["No mask needed: AI rebuilds fossil leaves from photo","Synthetic-only training rebuilds whole fossil leaf","Fossil leaf recovery: no visible mask, just RGB","AmodalDINO: rebuilding fossil leaves without a mask"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000664,"raw_usage":{"total_tokens":3123,"prompt_tokens":1128,"completion_tokens":1995,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":1929}},"tokens_in":744,"tokens_out":1995,"duration_ms":12768,"temperature":1.0,"reasoning_tokens":1929,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:16:48.057153+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Vision transformers for dense prediction,","cited_arxiv_id":null,"evidence_quote":"Supplies the DPT trunk that fuses four intermediate token blocks into the dense features read out by the four heads."},{"cited_title":"Amodal instance segmen- tation with KINS dataset,","cited_arxiv_id":null,"evidence_quote":"Defines the KINS dataset and its full/occluded mIoU protocol used for the public-benchmark comparison."},{"cited_title":"Semantic amodal segmen- tation,","cited_arxiv_id":null,"evidence_quote":"Defines the COCOA/COCOA-cls dataset and protocol used as the second public amodal benchmark."},{"cited_title":"Tversky loss function for image segmentation using 3D fully convolutional deep networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Tversky loss used in the completion heads to penalize over-growth into stone more than missed pixels."},{"cited_title":"Learnable Shape Prototypes with Occlusion-Geometry-Guided Injection for Amodal Instance Segmentation","cited_arxiv_id":"2605.24533","evidence_quote":"GRASP is the comparison baseline that, like this work, builds on a DINO backbone but keeps it frozen and consumes a visible mask."}],"review_version":1}