{"id":"1596ec88-7cf2-4375-a6a6-7f8ae9408a4d","arxiv_id":"2412.02336","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper introduces ADIW, a 564K-image pseudo-labeled dataset for relative amodal depth, and two fine-tuned models (Amodal-DAV2 and Amodal-DepthFM) that predict occluded-object depth from an image, observed depth, and amodal mask.","lead":"The authors build a large synthetic-style dataset for predicting the depth of hidden parts of objects in natural images, with the hidden depth labels produced by a pre-trained depth model. They then fine-tune two existing depth models, Depth Anything V2 and DepthFM, to make those predictions from a visible image, a visible depth map, and an amodal mask.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ADIW ground truth is Depth Anything V2's own output, so the reported SoTA measures self-consistency with the teacher rather than physical amodal depth.","rationale":"The reader's weakest assumption identifies exactly this issue, and I agree. The pipeline in Sec. 3.2 uses DA V2 to produce both Do and Db, then fits a global scale-and-shift on visible pixels. This assumes the depth relationship between the composite and background is a global affine transform; if DA V2's predictions for the hidden part are biased, the labels inherit that bias. Because the validation set is generated identically, the benchmark cannot distinguish correct amodal geometry from consistent reproduction of the teacher's bias. An independent synthetic test with known geometry is the decisive check. I also note the abstract's '69.5% improvement' does not match Table 1's 27.4% RMSE improvement, which is a secondary reporting inconsistency. The conditionality of the reader's verdict is appropriate: the dataset and method are useful engineering contributions, but the physical-accuracy claim remains unverified. Therefore I would keep the verdict unchanged.","tokens_in":13663,"tokens_out":5510,"duration_ms":62248,"concrete_test":"Build a synthetic benchmark with renderer-known ground-truth depth for occluded regions (e.g., Blender or Kubric scenes with objects occluded by other objects). Evaluate Amodal-DA V2-L, Amodal-DepthFM, and the DA V2 teacher (using the paper's Eq. 2 alignment on the un-occluded background) on the occluded-region depth error against the true renderer depths. If the fine-tuned models do not substantially reduce error relative to the teacher on these scenes, the ADIW numbers reflect self-consistency with DA V2 rather than accurate amodal geometry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that ADIW's ground-truth labels are generated by Depth Anything V2 itself, and the best model (Amodal-DA V2-L) is initialized from that same model. Equations 1-2 define the occluded-region depth as a scale-and-shift aligned version of DA V2's depth on the un-occluded background image; if DA V2's relative depth for hidden geometry is systematically wrong, every ADIW label is wrong in the same way. The validation split is produced by the same pipeline, so the reported 27.4% RMSE improvement (and the abstract's 69.5% accuracy claim) measure agreement with DA V2's priors, not physical amodal depth. The paper itself states that no device can capture ground truth for occluded regions (Sec. C), so the benchmark is self-consistent but not externally validated. This does not invalidate the dataset as a training resource, but it means the central claim of achieving amodal depth estimation 'in the wild' is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes relative amodal depth estimation \"in the wild\": given an RGB image, an observation depth map Do, and an amodal mask Ma, predict depth for the occluded part of the target object. The authors introduce ADIW, a 564K-pair dataset built by compositing an occluder over a background image taken from SAM/SA-1B segmentations; DA V2 (ViT-G) produces depth for both the composite and the background, and a per-image scale-and-shift fit on visible pixels (Eqs. 1-2) generates the occluded-region ground truth. Two models are fine-tuned from pre-trained depth models with minimal architectural changes: Amodal-DA V2, a DA V2 ViT with an extra zero-initialized guidance convolution and layer norm in the DPT head, trained with the scale-invariant log loss; and Amodal-DepthFM, a conditional flow-matching model extended with Do and Ma guidance channels, with scale-and-shift alignment applied at inference. On the ADIW validation split, the best model (Amodal-DA V2-L) reports RMSE 3.418 versus 4.712 for the retrained previous-SoTA baseline Jo et al. (a 27.4% relative improvement), and the abstract claims a 69.5% accuracy improvement over the previous SoTA. Ablations show that the guidance channels, whole-object supervision, and (for DepthFM) inference-time alignment each help. The paper also shows qualitative zero-shot results and 3D mesh reconstructions of occluded objects.","tokens_in":13804,"tokens_out":18501,"duration_ms":191202,"significance":"The main value of the paper is in dataset construction and model engineering. ADIW appears to be the first large-scale training resource for relative amodal depth, the compositing pipeline is genuinely scalable, the two model adaptations are simple and clearly described, and the ablations in Tables 2-3 are informative. The paper is also transparent that no existing sensor can directly measure occluded-region depth (Secs. 1 and 3.2). However, the significance is conditional: the ADIW validation labels are generated by DA V2 itself (Sec. 3.2, Eqs. 1-2), the best model is initialized from DA V2 (Sec. 4.2), and the reported state of the art therefore currently demonstrates accurate reproduction of DA V2's hidden-geometry predictions rather than physical amodal depth. The contribution would become a clear accept with external validation against true occluded geometry (for example, synthetic benchmarks with rendered ground truth or a real-sensor split), or with the claims explicitly reframed as teacher-agreement results. As written, the headline numbers in the abstract and Sec. 4.3 overstate what has been established.","major_comments":[{"comment":"The ADIW ground truth for occluded regions is defined as DA V2's relative depth on the un-occluded background image, aligned to the observation depth with a per-image scale-and-shift fit (Eqs. 1-2), and the 4K-sample validation split is produced by the same pipeline. The best model, Amodal-DA V2-L, is initialized from DA V2 and fine-tuned on these labels (Sec. 4.2). The headline numbers in Table 1 (the 27.4% RMSE improvement over Jo et al. retrained on ADIW, and the abstract's 69.5% accuracy claim) therefore measure how faithfully the fine-tuned model reproduces DA V2's own hidden-geometry estimates, and a DA V2-initialized model has an inherent advantage over baselines that must learn the teacher's prior from the labels alone. The paper's own statements that no device can capture occluded-region ground truth (Secs. 1 and 3.2) make this a load-bearing limitation rather than a side remark. This does not invalidate ADIW as a training resource, but the claim of amodal depth estimation \"in the wild\" is not yet established. Please reframe the evaluation and SoTA claims as measuring agreement with the teacher prior, and add external validation of physical accuracy, for example zero-shot evaluation on Amodal-SynthDrive and Amodal-3D-FRONT (which have rendered ground-truth depth behind occluders) or on a split where a real RGB-D sensor measures the background depth before a physical occluder is placed.","section":"Secs. 1, 3.2, 4.3; Eqs. (1)-(2)"},{"comment":"The evaluation protocol does not specify how predicted relative depth maps are brought into the scale of the ground truth before RMSE, log10, and delta are computed. The ground truth is defined in the canonical scale of Do (normalized to [0,1] and then scale-shifted per image, Eqs. 1-2), while the baselines marked with a double dagger are explicitly scale-shift aligned to the observation depth at inference and Amodal-DA V2 is not. Since all three metrics are scale-sensitive, the absolute numbers in Table 1 cannot be independently reproduced without stating the normalization applied to each method's raw output (for example, whether predictions are re-normalized to [0,1], fitted to Do, or used as-is). Please make the protocol explicit and show that the reported differences are not an artifact of scale matching; this is needed for the 27.4% SoTA claim to be verifiable.","section":"Sec. 4.1 and Table 1"},{"comment":"The abstract's 69.5% accuracy claim is not consistent with the body's 27.4% RMSE claim. The 27.4% figure is reproducible from Table 1 against the aligned retrained previous-SoTA row (Jo et al. with double-dagger alignment, RMSE 4.712 versus 3.418). The 69.5% figure is not: relative to the aligned Jo et al. row, the delta accuracy improves by about 5.6% (89.177 to 94.179), whereas the 69.5% value matches the relative delta improvement against the unaligned Jo et al. row (55.545 to 94.179), which is the weaker configuration the table itself marks as not the recommended comparison. Please state explicitly which baseline and which metric the abstract's headline number refers to, and recompute it against the same configuration used for the SoTA comparison in Sec. 4.3.","section":"Abstract and Sec. 4.3"}],"minor_comments":[{"comment":"The scale-invariant loss Lsi is written with constants alpha and lambda, but alpha is never defined; with lambda = 0.85 the loss is only partially scale-invariant, which is directly relevant to the scale-normalization question raised above, so please define all constants and state the resulting scale behavior of the training target.","section":"Sec. 3.3.1"},{"comment":"The flow-matching objective is written as min over theta of the expectation of the norm of v_theta(t, phi_t(x0)) minus (x1 - x0), but the norm subscript is missing, the marginalization over p(x0) is not defined, and the notation mixes the conditioning code z with phi_t(x|z); please clean up the notation.","section":"Sec. 3.3.2"},{"comment":"The phrase \"previous SoTA amodal depth model\" is applied to Jo et al. [16] even though ADIW is a new benchmark and the comparison is run on a model retrained on ADIW; please reword to make clear that the baseline is the best existing amodal depth architecture retrained on the new dataset.","section":"Sec. 4.3 and Table 1"},{"comment":"The model name alternates between \"Amodal-DAV2\" and \"Amodal-DA V2\"; please use one consistent form throughout the paper.","section":"Abstract and throughout"},{"comment":"References [48] and [49] are the same paper (Zhan et al., \"Self-supervised scene de-occlusion,\" CVPR 2020) and are cited as two distinct entries; please merge them.","section":"References"},{"comment":"The metrics \"log 10 error\" and \"accuracy under the threshold delta\" are never given formulas; since these are scale-sensitive for relative depth, please define them (for example, mean absolute log10 error, and the fraction of pixels satisfying max(y/y_hat, y_hat/y) < 1.25).","section":"Sec. 4.1"},{"comment":"The limitation section discusses dependence on the amodal mask and a decline in detail-capturing ability, but it does not mention that the benchmark ground truth is teacher-generated; adding this caveat would help future users interpret ADIW scores correctly.","section":"Appendix C"},{"comment":"The dataset is described as \"real-world data\" in one place and \"realistic synthetically generated datasets\" in another; since the compositing step is synthetic, the wording should be consistent to avoid confusion about what is real versus what is composed.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the reader's CONDITIONAL verdict and my own assessment converge on one central issue: the ADIW benchmark labels are DA V2's own outputs, the top model is initialized from DA V2, and the validation split shares the label pipeline. I regard this as fixable within a major revision if the authors either add external validation (for example, synthetic amodal benchmarks with known geometry, or a real-sensor background-depth split) or explicitly reframe all SoTA claims as teacher-agreement results. The abstract's 69.5% accuracy claim versus the body's 27.4% RMSE claim should also be reconciled during revision; the 69.5% figure appears to be computed against the unaligned, weaker baseline configuration. If the circularity is not addressed, the paper would overclaim amodal depth estimation \"in the wild\"; with an external check and honest reframing, it is a solid dataset-and-models contribution that fits the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper gives the field a large-scale real-image dataset for amodal relative depth and two fine-tuned models that beat the baselines it retrains. That is the real contribution. The relative depth formulation is a sensible break from the metric-depth framing in prior synthetic-only work, and the ADIW pipeline (compositing + scale-shift alignment) is clearly described and reproducible. The ablations are informative: guidance channels and object-level supervision both matter, and the generative model's diversity is a nice touch.\n\nThe soft spot is the one you already see. The ground-truth occluded depths are Depth Anything V2's own output, aligned from the un-occluded background image via Eqs. 1-2, and the best model is initialized from the same model family. The validation split is generated by that same pipeline, so the reported 27.4% RMSE improvement (and the abstract's 69.5% accuracy claim) largely measure how well Amodal-DA V2-L reproduces DA V2's priors on occluded regions. This is not a definitional identity, but it is load-bearing. The paper honestly states that no device can capture ground truth for occluded parts (Sec. C), so the benchmark is internally consistent but not externally validated.\n\nThat said, the circularity does not kill the paper. As a training resource and a task formulation, ADIW is useful even if the labels carry DA V2's systematic biases. The models are simple but effective, and the relative depth framing is likely to generalize better than the synthetic metric baselines. What I cannot accept is the claim that the method estimates physical amodal depth 'in the wild' with SoTA accuracy. That requires external validation, for example on synthetic scenes where the hidden geometry is known, or a small hand-verified real set with a different sensing modality.\n\nBottom line: send it to peer review. A good referee should ask for external validation before accepting the quantitative claims, but the dataset and frameworks are worth publishing even with the caveats. I'd cite the dataset if I worked in this area.\n\nThis is a solid engineering paper with an honest limitations section, not a settled measurement paper.","headline":"Useful dataset and two working models for amodal relative depth, but the headline numbers measure agreement with the teacher, not physical depth.","tokens_in":14408,"tokens_out":1339,"would_cite":true,"duration_ms":16807,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that amodal depth — the depth of object parts hidden behind occluders — can be learned from real images by reposing the task as relative depth and training on a large composited dataset with Depth Anything V2 as teacher.","keywords":["amodal depth estimation","relative depth prediction","ADIW dataset","occluded geometry","scale-and-shift alignment","flow matching depth","Depth Anything V2","depth inpainting"],"falsifier":"Take real scenes with a known object partially hidden behind another object, measure the true depth of the hidden surface after removing the occluder, for example with a depth sensor or multi-view reconstruction, and compare those measurements with both the model predictions on the occluded image and ADIW's assembled labels. If the physical hidden depths disagree systematically with the model or with the aligned Depth Anything V2 labels, the reported accuracy would be measuring agreement with a teacher prior rather than true amodal geometry.","tokens_in":13425,"feed_emoji":"📏","tokens_out":5978,"duration_ms":61253,"temperature":0.7,"pith_summary":"The paper tries to establish that the depth of occluded, invisible parts of objects can be predicted from a single natural image, and that this amodal depth task is best posed as relative depth rather than metric depth. To support that claim it introduces ADIW, a 564K-image dataset built by compositing segmented objects onto natural backgrounds and labeling them with depth from a large pre-trained relative-depth model, then aligning scales so the hidden geometry is consistent with visible geometry. On this dataset, the deterministic Amodal-DA V2 model and the generative Amodal-DepthFM model both beat prior amodal depth methods and inpainting-based alternatives, with the best model reporting a 27.4% RMSE improvement over the previous state of the art. A sympathetic reader would care because, if true, it means geometry behind occluders is learnable from real-world image statistics rather than requiring hand-built synthetic scenes.","feed_headline":"Model estimates depth of hidden object parts, cutting error by 27.4%","feed_subtitle":"New ADIW dataset trains relative-depth models on composited real images so occluded geometry becomes predictable.","key_machinery":"The scale-and-shift alignment is the load-bearing mechanism: for each composited training pair, Eqs. 1-2 fit the background image's depth to the observed image's visible depth with least-squares scale $s$ and shift $t$, producing $D_{\\mathrm{aligned}} = sD_b + t$ as ground truth. This transfers the pre-trained model's depth for the hidden part into the composite. The second mechanism is the guidance-channel modification: a zero-initialized parallel convolution adds the observation depth map $D_o$ and amodal mask $M_a$ to the encoder, so fine-tuning starts from pre-trained weights and the model learns to condition on those signals. Together these convert a monocular depth foundation model into an amodal depth predictor with minimal architecture change.","core_discovery":"The central claim is that relative amodal depth can be learned in the wild: given an image, its observed depth map, and an amodal mask marking where an object extends behind an occluder, a model fine-tuned from Depth Anything V2 can output a depth map that includes the occluded region, and this transfers to real images better than previous metric-depth amodal models. The dataset construction is the enabling move: rather than measuring hidden depth directly, which no sensor can do at scale, the pipeline uses the depth a strong relative-depth model assigns to an un-occluded image of the same object, aligned to the visible depth by a global scale and shift. The paper further claims that both the deterministic and generative adaptations work, that object-level supervision helps, and that the guidance channels carrying the observed depth and the amodal mask account for most of the performance gain.","pith_inferences":["Beyond the paper, the same compositing-and-alignment pipeline could be reused to train amodal models for surface normals, segmentation, or RGB inpainting by swapping the teacher model.","Because ADIW labels are inherited from Depth Anything V2, the benchmark's improvements are partly a measure of consistency with that model's geometric prior; a physical-ground-truth test could change the ranking.","The framework's real-world value depends on upstream amodal segmentation quality, since the limitation section notes that inaccurate or ambiguous masks propagate errors into depth predictions.","A direct evaluation with depth sensors on partially occluded real objects would separate learned geometric priors from alignment artifacts; the paper does not provide such an evaluation."],"forward_implications":["If the central claim holds, relative-depth amodal estimation can be trained from real images at scale without any physical sensor for occluded depth.","The reported results imply that the ADIW benchmark separates methods on occluded-region accuracy, with Amodal-DA V2-L reducing RMSE by 27.4% over the previous state of the art.","Ablations imply that both the observed-depth and amodal-mask guidance channels are necessary, since removing either degrades accuracy.","The generative variant implies that multiple plausible hidden geometries can be sampled for one scene, not just a single deterministic prediction.","The comparison with inpainting-based baselines implies that direct depth regression is more reliable than inpainting-then-depth-estimation pipelines for occluded geometry."],"supporting_citations":[{"why":"Supplies the pre-trained Depth Anything V2 model used both to generate ADIW depth labels and as the backbone for Amodal-DA V2.","marker":"[46]"},{"why":"Provides the whole-object dataset construction method and compositing technique used to build training pairs, and also serves as an inpainting baseline.","marker":"[27]"},{"why":"Supplies the scale-and-shift alignment algorithm used in Eqs. 1-2 to make the composited depth labels consistent.","marker":"[33]"},{"why":"Provides the DepthFM flow-matching model that is adapted into Amodal-DepthFM.","marker":"[13]"},{"why":"Generates the segmentation masks from SA-1B that seed the whole-object dataset.","marker":"[20]"},{"why":"The prior amodal depth method on synthetic data that is retrained and compared against as the main baseline.","marker":"[16]"},{"why":"An inpainting-and-depth baseline used to show that direct amodal depth regression avoids cascade errors.","marker":"[10]"},{"why":"Supplies the earlier Depth Anything model and the insight that relative depth estimators generalize better than metric ones.","marker":"[45]"}],"fun_headline_variants":["New dataset boosts amodal depth accuracy by 69.5%","AI predicts depth behind occluders in real-world scenes","Relative depth estimation sees occluded parts in the wild","Flow matching model infers hidden object depth","Amodal depth in the wild: from visible cues to invisible geometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the depth a large pre-trained model assigns to an object in an un-occluded background image is the true depth of that object's hidden geometry in the composite, after a global scale-and-shift fit on visible pixels.","fun_headline_variants_meta":{"raw":{"variants":["New dataset boosts amodal depth accuracy by 69.5%","AI predicts depth behind occluders in real-world scenes","Relative depth estimation sees occluded parts in the wild","Flow matching model infers hidden object depth","Amodal depth in the wild: from visible cues to invisible geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3731,"prompt_tokens":980,"completion_tokens":2751,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2670}},"tokens_in":596,"tokens_out":2751,"duration_ms":22956,"temperature":1.0,"reasoning_tokens":2670,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:34:38.448742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take real scenes with a known object partially hidden behind another object, measure the true depth of the hidden surface after removing the occluder, for example with a depth sensor or multi-view reconstruction, and compare those measurements with both the model predictions on the occluded image and ADIW's assembled labels. If the physical hidden depths disagree systematically with the model or with the aligned Depth Anything V2 labels, the reported accuracy would be measuring agreement with a teacher prior rather than true amodal geometry.","supporting_citations":[{"cited_title":"pix2gestalt: Amodal segmentation by synthesizing wholes","cited_arxiv_id":null,"evidence_quote":"Provides the whole-object dataset construction method and compositing technique used to build training pairs, and also serves as an inpainting baseline."},{"cited_title":"Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer","cited_arxiv_id":null,"evidence_quote":"Supplies the scale-and-shift alignment algorithm used in Eqs. 1-2 to make the composited depth labels consistent."},{"cited_title":"Occlusion- aware amodal depth estimation for enhancing 3d reconstruc- tion from a single image","cited_arxiv_id":null,"evidence_quote":"The prior amodal depth method on synthetic data that is retrained and compared against as the main baseline."}],"review_version":1}