{"id":"ac0b2d6f-8233-44b6-a44d-ca03a9039cec","arxiv_id":"2505.10565","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Prior Depth Anything unifies depth completion, super-resolution, and inpainting by aligning a monocular prediction to partial metric priors pixel by pixel, then refining with a conditioned network.","lead":"This paper presents a single model that merges coarse but accurate depth measurements with a complete but relative depth map predicted from the image, producing dense metric depth for arbitrary scenes. The model handles sparse points, low-resolution maps, missing areas, and mixtures of these input types without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pre-fill's local surface assumption (Eqs. 2/4) breaks when KNN support spans depth discontinuities — most likely in the 100-point 'Extreme' setting — and the paper never quantifies how often this occurs or how well refinement recovers.","rationale":"The paper proposes a coarse-to-fine fusion: per-pixel KNN affine alignment (Eqs. 1-4) pre-fills arbitrary missing patterns in a metric prior with structure borrowed from a frozen MDE, and a conditioned MDE (metric plus geometry conditions, zero-initialized) refines the result. For the central claim, three things must be true: (i) the pre-fill produces a useful intermediate for all prior forms; (ii) the refinement converts it into accurate dense metric depth; (iii) the comparative claims hold. The least-supported of these is (i) in the regimes the paper most emphasizes (extreme sparsity, discontinuity-rich scenes), because the local-affine assumption that a single (s,t) is valid for the query and its support silently fails when the support spans multiple surfaces, and the paper provides no error decomposition by support dispersion. The end-to-end numbers (e.g., NYU Extreme 1.96-2.01 AbsRel) are strong, and the refinement demonstrably cuts pre-fill error roughly in half (Table 6 vs. Tables 2-5), so the concern is not that the method fails outright; it is that the boundary of the 'any prior' claim is unquantified. The proposed test resolves this directly. Secondary observations, none changing the core verdict: (a) the abstract's 'matching or even surpassing' is contradicted by Table 4, where task-specific PromptDA beats PriorDA on six of seven super-resolution benchmarks, including ARKitScenes, and the body concedes only 'comparable' performance; (b) the scale de-normalization in Sec. 3.3 is never specified, and no code is shipped, so reimplementation is not currently possible; (c) Eq. 1 sets valid-pixel pre-fill values to the relative prediction, contradicting 'inherits all the valid pixels', presumably a typo; (d) the manuscript contains a severe artifact: roughly four pages of Depth Anything V2's Section 5, repeated about seven times, are inserted mid-paper; this does not bear on correctness but must be cleaned. I agree with the reader's weakest_assumption: the KNN surface-straddling failure is the load-bearing risk. Because the risk is testable and the verdict was already CONDITIONAL, I recommend UNCHANGED.","tokens_in":26161,"tokens_out":16113,"duration_ms":144596,"concrete_test":"Recompute the pre-filled and final outputs (Eqs. 1-4, K=5) on NYUv2 and KITTI with the 100-point Extreme prior, bucketing missing pixels by the depth dispersion (max-min) of their K nearest valid support points. If final AbsRel on the high-dispersion bucket (support straddling a depth edge) is not significantly worse than on the low-dispersion bucket, the conditioned MDE recovers and the concern is resolved; if it is worse, the 'any prior' claim needs qualification for discontinuity-rich and extremely sparse inputs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is unified zero-shot dense metric depth from any prior. Its load-bearing mechanism is the coarse pre-fill: for each missing pixel, fit one scale s and shift t mapping the frozen MDE's relative prediction to the K nearest valid prior pixels (Eqs. 2/4), then fill the query as s*D_pred+t. This assumes the query and its KNN support lie on one locally smooth surface so a single affine map is valid. The assumption breaks exactly where the method is advertised to excel: at depth discontinuities and in the 100-point 'Extreme' setting, where nearest valid points can be far apart and straddle two surfaces. The fitted (s,t) is then a blend; a foreground object whose nearest neighbors lie on the background is pre-filled at the wrong metric scale, and Eq. 4's inverse-distance weighting mitigates but cannot remove this systematic bias. MDE outputs are affine-invariant inverse depth, so an affine map to metric depth is exact only when the support's depth range is small, which is precisely the condition violated here. The paper never quantifies the failure: no measurement of how often KNN support spans a depth edge as a function of sparsity, and no breakdown of pre-fill or final error on such pixels. Ablations (Tables 6, 12) report aggregates that wash out this failure mode. Related gaps: Eq. 1 appears to be a typo (valid pixels are set to the prediction, contradicting 'inherits all the valid pixels'), and the scale de-normalization in Sec. 3.3 is unspecified, so the pipeline is not currently reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Prior Depth Anything, a single framework for dense metric depth estimation from an RGB image plus an arbitrary incomplete metric depth prior (sparse points, low-resolution depth, missing areas, or mixtures). The method first uses a frozen MDE model to obtain a relative depth prediction, pre-fills missing prior pixels via per-pixel affine alignment to the K nearest valid prior points (Eqs. 2-4), and then refines the pre-filled map with a conditioned MDE model trained on synthetic data. The paper reports zero-shot results on 7 real-world datasets for depth completion, super-resolution, and inpainting, and demonstrates an application to refining VGGT predictions. The core design is clearly described and the experimental study is broad, but the paper's central claim of matching or surpassing task-specific methods is not uniformly supported by its own tables.","tokens_in":26496,"tokens_out":7308,"duration_ms":63175,"significance":"If accepted with appropriately qualified claims, the central idea of converting all prior patterns into a shared pre-filled domain and then learning to refine it is useful and practical. Strengths include the systematic zero-shot evaluation across 7 datasets and 9 prior patterns, the clean ablations (Tables 6-8, 12) isolating the contribution of alignment, re-weighting, and the two conditioning streams, the test-time MDE switching (Table 9), and the efficiency advantage over diffusion-based baselines (Table 10). The main weaknesses are an overbroad claim in the abstract and conclusion, an unspecified de-normalization step, and a lack of failure-mode analysis for the KNN-based pre-filling at depth discontinuities and extreme sparsity. These issues are addressable, but they require more than local edits.","major_comments":[{"comment":"The central claim that the method is 'matching or even surpassing previous task-specific methods' is contradicted by the paper's own tables. On ARKitScenes super-resolution (Table 4), PromptDA achieves AbsRel 1.34 / RMSE 0.0347 while PriorDA (DAv2-B+ViT-B) achieves 1.94 / 0.0404; in depth completion (Table 3), Marigold-DC outperforms all PriorDA variants on ScanNet LiDAR (1.76 vs 2.15/2.19) and Omni-DC outperforms them on KITTI SfM (3.34 vs 3.35/3.73). The abstract and conclusion therefore overstate the result. Please qualify the claim to 'competitive or superior in most settings' and discuss where task-specific methods retain an advantage.","section":"Abstract; Section 4.3, Table 4"},{"comment":"The coarse alignment assumes that the K nearest valid points and the query pixel lie on a locally smooth surface, so that a single affine map (s,t) is valid. When the prior is extremely sparse (100 points, Table 3 'Extreme') or when the KNN support spans a depth discontinuity, the fitted affine map is a blend of two surfaces and the pre-filled value is systematically biased. The paper does not quantify how often this occurs, nor how well the refinement stage recovers from it; the aggregate ablations in Tables 6 and 12 can mask this failure mode. Please add an analysis, for example pre-fill and final error as a function of distance to the nearest depth boundary or of local valid-point density, and discuss the impact on the 'Extreme' setting.","section":"Section 3.2, Eq. (4); Table 3 'Extreme'; Tables 6 and 12"},{"comment":"The de-normalization from the normalized metric and geometry conditions back to the ground-truth scale is not specified. The text states 'we apply the de-normalization transformation to convert the output into the ground truth scale' but gives no equation or procedure. This is necessary for reproducibility and for understanding how the metric scale is recovered after the [0,1] normalization. Please state the exact transformation (for example, re-scaling by per-image prior statistics) and indicate where it is applied in the pipeline.","section":"Section 3.3 and 3.4"}],"minor_comments":[{"comment":"Equation (1) sets the pre-filled map at valid pixels to Dpred(x,y), which contradicts the text 'inherits all the valid pixels in Dprior' and the surrounding description; it should be Dprior(x,y).","section":"Section 3.2, Eq. (1)"},{"comment":"The k-value is set to 5, but Section 3.2 calls the alignment 'parameter-free'; since K is a hyperparameter, consider using 'with no learned parameters' or 'unlearned' to avoid confusion.","section":"Section 3.4"},{"comment":"The denominator is written as ||(x_hat,y_hat)-(x_k,y_k)||_2, which can be read as squared Euclidean distance, while the text says 'based on its distance'; please clarify whether the weight is 1/d or 1/d^2.","section":"Eq. (4)"},{"comment":"The entry 'Depth Pro+ViT-B2.01' appears to be missing a space and is difficult to read; please check the table formatting.","section":"Table 2"},{"comment":"The sentence 'On these two benchmarks, our method achieves leading performance compared to other zero-shot methods' is not supported by Table 4, where PromptDA has lower AbsRel on ARKitScenes; please align the text with the reported data.","section":"Section 4.3, text"},{"comment":"The manuscript contains a long excerpt from the Depth Anything V2 paper (beginning with '5 Depth Anything V2' after the references), which is not part of this paper; it should be removed before submission.","section":"Appendix (after References)"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically substantial and the experimental study is unusually broad for this line of work. The main risks are the overbroad headline claim, which needs to be qualified, and the underspecified de-normalization step. I also noticed that the manuscript text contains a large copied block from a different paper (Depth Anything V2) in the appendix; this should be cleaned up before resubmission. I do not see evidence of circular benchmarking: the per-pixel alignment is computed on the fly from the input prior, not fitted to the evaluation datasets."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper deserves a serious referee, but the current version is not ready for publication as-is. The core idea is real: a single zero-shot model that handles sparse points, low-resolution maps, missing areas, and mixtures of these, using a pixel-level affine pre-fill followed by a conditioned MDE refinement. The ablations (Tables 6–8) support the design choices, and the mixed-prior results are genuinely strong. The \"any prior\" framing is not just marketing; the method does something none of the cited task-specific works do.\n\nWhat is actually new: the distance-aware, per-pixel affine alignment of a frozen MDE prediction to the metric prior (Eqs. 2 and 4) is a clean idea, and the zero-initialized conditioning layers that preserve the pretrained MDE's geometry knowledge are well motivated. The test-time MDE swapping (Table 9) is a nice bonus. Credit where it is due: the paper is clearly written, the method is specified in enough detail that a competent grad student could reimplement most of it, and the evaluation covers seven real datasets with nine prior patterns. That is a solid experimental footprint.\n\nNow the soft spots, in proportion to how soft they are. The blanket claim of \"matching or even surpassing previous task-specific methods\" is contradicted by the paper's own Table 4: PromptDA gets 1.34 AbsRel on ARKitScenes versus 1.94 for the best PriorDA variant, and that is not a minor gap. The claim should be softened to \"competitive overall, best on mixed priors.\" Second, the stress-test concern about KNN support straddling depth discontinuities is legitimate but not fatal. The distance weighting and the learned refinement clearly recover a lot (pre-fill 2.42 → final 2.01 for sparse points on NYUv2), but the paper never quantifies how often the pre-fill fails near edges or at the 100-point extreme. A reviewer should ask for exactly that breakdown. Third, Eq. 1 is a typo: it sets valid pixels to D_pred, contradicting the text that says the pre-filled map inherits valid pixels from D_prior; that needs fixing. Fourth, the scale de-normalization in Sec. 3.3 is underspecified — \"apply the de-normalization transformation\" without saying what the transform depends on leaves the pipeline not fully reproducible, and there is no code. Finally, the duplicated Depth Anything V2 blocks in the full text are a severe artifact; this looks like a LaTeX submission error, but it must be cleaned before any referee sees it.\n\nWho this is for: researchers working on depth completion, super-resolution, inpainting, or any sensor-fusion application. The paper is useful as a capable unified baseline and as a source of the pre-fill idea. My recommendation: offer peer review. A good referee will ask for corrected claims, the edge-failure analysis, and a reproducible de-normalization, but the core contribution is sound and the experimental work is honest enough to build on.","headline":"A genuinely unified any-prior depth model with a clean coarse-to-fine design, but the current manuscript overstates its benchmark wins and needs a thorough edit before it is referee-ready.","tokens_in":27042,"tokens_out":3614,"would_cite":true,"duration_ms":36962,"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":"One depth model now takes any partial depth measurement—sparse, blurry, or hole-filled—and returns dense metric depth, without task-specific training.","keywords":["depth completion","depth super-resolution","depth inpainting","metric depth prior","zero-shot generalization","monocular depth estimation","pixel-level affine alignment","conditioned MDE"],"falsifier":"A controlled experiment on a depth-completion benchmark like KITTI or NYUv2 that densely samples occlusion boundaries, then provides a decreasing sequence of valid prior points (e.g., 500, 100, 50, 20 pixels) and measures whether the final refined output's error rises sharply when the pre-filled values cross depth edges. If the error increases in proportion to the number of pixels whose K nearest neighbors cross a depth discontinuity, the core alignment assumption is doing the work; if the refinement stage fully recovers, the strong claim of robustness to any prior holds. A second check is to feed a prior that is deliberately non-affine within each local neighborhood (e.g., valid points sampled from two different walls at different depths) and see whether the model propagates the affine fit error into the filled region.","tokens_in":25961,"feed_emoji":"🗺️","tokens_out":3839,"duration_ms":31252,"temperature":0.7,"pith_summary":"The paper sets out to prove that a single framework can absorb any incomplete metric depth prior—sparse LiDAR or SfM points, low-resolution sensor maps, missing regions, or any mixture of these—and, guided by an RGB image, output dense, detailed, metrically accurate depth in a zero-shot fashion. The authors argue that previous methods each handle only one pattern of incomplete depth, so they fail on mixed or unseen priors. Prior Depth Anything is their proposed unification: first explicitly align a frozen monocular depth prediction to the metric prior at the pixel level using per-pixel least-squares scale-and-shift fits, then refine the result with a depth-conditioned monocular model that merges metric values with predicted geometry. Across seven real-world datasets the method claims performance matching or beating task-specific systems for depth completion, super-resolution, and inpainting, while remaining the only approach that handles all three at once.","feed_headline":"One model completes any partial depth map","feed_subtitle":"Sparse LiDAR, low-res ToF, or hole-filled depth maps all become dense metric depth with a single zero-shot framework.","key_machinery":"The coarse metric alignment step is the load-bearing mechanism: for each missing pixel it finds the K nearest valid prior points, fits an optimal scale s and shift t by weighted least squares (weighted by inverse squared distance) between the frozen MDE prediction and the metric prior at those points, then fills the missing pixel as s·D_pred + t. This transforms every prior pattern into the same intermediate domain, preserving the prior's metric values while inheriting the prediction's fine geometry. The fine structure refinement then feeds the normalized pre-filled prior and normalized prediction into a conditioned MDE model through zero-initialized convolutional layers, so the model learns to correct noisy millimeter-level priors while keeping the metric accuracy. Test-time model switching is enabled by normalizing the prediction before conditioning and applying the inverse normalization after the conditioned model produces its output.","core_discovery":"The paper discovers that a frozen monocular depth estimation model's relative prediction can act as a universal geometric scaffold for any incomplete metric prior, and that pixel-wise affine alignment turns that scaffold into a dense metric pre-fill which a conditioned MDE model can then refine. The central claim is that coarse-to-fine integration—pre-fill by per-pixel scale-shift fits to the K nearest valid prior points, then refine with a conditioned MDE that sees both the pre-filled metric prior and the normalized relative prediction—closes the gap between any prior pattern and a dense metric depth map. The authors show the pre-fill alone narrows the domain gap among prior types, so a model trained on just sparse points generalizes to low-resolution and missing-area priors. The refinement stage is what corrects noisy measurements and improves the final output over the pre-fill. The result is a single model that claims state-of-the-art or competitive zero-shot results on depth completion, super-resolution, and inpainting across seven datasets, and that uniquely handles mixed priors.","pith_inferences":["A natural extension the paper does not explore is using the same coarse-to-fine integration for video or multi-frame depth, where temporal consistency between pre-fills could smooth noisy sensor streams and correct flicker in the refined output.","The pre-fill's reliance on local affine fits suggests a harder test than any single dataset: scenes with sharp occlusion boundaries and very sparse priors (a few hundred points) may expose whether the refinement stage can recover structural errors, not just sensor noise. A benchmark comparing pre-fill accuracy vs refinement recovery on occlusion-heavy scenes would quantify this.","The conditioned MDE is trained on synthetic priors sampled from Hypersim and vKITTI, so a distribution shift to real-sensor noise patterns (e.g., ToF multipath interference or LiDAR beam divergence) is where the claimed generalization is least certain; testing on real captured low-resolution and partial-depth cameras beyond ARKitScenes and RGB-D-D would be a direct stress test."],"forward_implications":["A single prior-based depth model can replace three separate specialist pipelines (completion, super-resolution, inpainting) in applications like AR/VR scene understanding, 3D reconstruction, and robotic navigation. If the central claim is correct, those applications no longer need to know which depth sensor or which failure mode produced the partial measurement.","Test-time swapping of the frozen MDE model (e.g., Depth Anything v2 ViT-G or Depth Pro for a stronger one) improves accuracy without retraining, while smaller components keep the pipeline efficient, so users can trade accuracy for latency per deployment.","Real-world depth measurements from noisy sensors can be refined with the metric prior preserved; the paper shows the model can correct blurred edges in widely used benchmark ground truths, meaning downstream tasks like 3D Gaussian editing and SLAM could consume cleaner dense depth maps.","The coarse metric alignment's pattern-agnostic pre-fill is what drives generalization, so the framework should absorb novel prior patterns not in the training set as long as the frozen MDE supplies reasonable relative geometry for the scene.","Because the conditioning is on normalized prior and prediction, the framework is compatible with any future relative-depth foundation model, so it improves automatically as MDE models improve."],"supporting_citations":[{"why":"Supplies the frozen and conditioned MDE backbones (Depth Anything v2 ViT-B/S) whose relative depth prediction provides the geometric scaffold for both alignment and refinement.","marker":"[56]"},{"why":"Depth Pro provides an alternative frozen MDE whose prediction is used as geometry scaffold and whose metric-depth capability is the baseline it outperforms; also supplies the 'Depth Pro+ViT-B' variant.","marker":"[4]"},{"why":"Omni-DC is the strongest zero-shot depth completion baseline whose per-pattern specialist design and multiresolution depth integration the method must beat on sparse/completion and mixed-prior settings.","marker":"[66]"},{"why":"Marigold-DC is the diffusion-based depth completion baseline whose prediction-guided refinement the method must beat on sparse-prior tasks.","marker":"[47]"},{"why":"PromptDA defines the depth super-resolution baseline that conditions a depth foundation model on low-resolution prompts; its training at high resolution is also acknowledged as a limitation the paper compares against.","marker":"[29]"},{"why":"DepthLab is the depth inpainting baseline using interpolation plus a diffusion refiner, and the main comparison for missing-area priors.","marker":"[30]"},{"why":"ZoeDepth contributes the scale-invariant log loss used as the learning objective for the conditioned MDE refinement stage.","marker":"[3]"},{"why":"Hypersim (plus vKITTI [6]) provides the precise synthetic training data with which the method constructs diverse synthetic priors and trains the conditioned MDE to rectify measurement noise.","marker":"[36]"}],"fun_headline_variants":["Any sparse prior becomes dense metric depth","One framework unifies all depth priors","Zero-shot depth from any partial input","Coarse-to-fine fusion of any depth prior","Pre-fill and refine: any depth prior works"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pre-fill assumes that the K nearest valid prior points to any missing pixel lie on the same local surface as that pixel, so a single scale and shift can convert the relative prediction into metric depth there; when the prior is extremely sparse or the scene has depth discontinuities, the nearest neighbors can straddle multiple surfaces and the filled value is wrong before the refinement stage ever sees it.","fun_headline_variants_meta":{"raw":{"variants":["Any sparse prior becomes dense metric depth","One framework unifies all depth priors","Zero-shot depth from any partial input","Coarse-to-fine fusion of any depth prior","Pre-fill and refine: any depth prior works"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000125,"raw_usage":{"total_tokens":1359,"prompt_tokens":947,"completion_tokens":412,"prompt_tokens_details":{"cached_tokens":896},"prompt_cache_hit_tokens":896,"prompt_cache_miss_tokens":51,"completion_tokens_details":{"reasoning_tokens":344}},"tokens_in":51,"tokens_out":412,"duration_ms":267468,"temperature":1.0,"reasoning_tokens":344,"cache_read_input_tokens":896,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:06:50.460668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment on a depth-completion benchmark like KITTI or NYUv2 that densely samples occlusion boundaries, then provides a decreasing sequence of valid prior points (e.g., 500, 100, 50, 20 pixels) and measures whether the final refined output's error rises sharply when the pre-filled values cross depth edges. If the error increases in proportion to the number of pixels whose K nearest neighbors cross a depth discontinuity, the core alignment assumption is doing the work; if the refinement stage fully recovers, the strong claim of robustness to any prior holds. A second check is to feed a prior that is deliberately non-affine within each local neighborhood (e.g., valid points sampled from two different walls at different depths) and see whether the model propagates the affine fit error into the filled region.","supporting_citations":[{"cited_title":"Depth any- thing v2","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen and conditioned MDE backbones (Depth Anything v2 ViT-B/S) whose relative depth prediction provides the geometric scaffold for both alignment and refinement."},{"cited_title":"Depth pro: Sharp monocular metric depth in less than a second","cited_arxiv_id":null,"evidence_quote":"Depth Pro provides an alternative frozen MDE whose prediction is used as geometry scaffold and whose metric-depth capability is the baseline it outperforms; also supplies the 'Depth Pro+ViT-B' variant."},{"cited_title":"Prompting depth anything for 4k resolution accurate metric depth estimation","cited_arxiv_id":null,"evidence_quote":"PromptDA defines the depth super-resolution baseline that conditions a depth foundation model on low-resolution prompts; its training at high resolution is also acknowledged as a limitation the paper compares against."},{"cited_title":"Hypersim: A photorealistic syn- thetic dataset for holistic indoor scene understanding","cited_arxiv_id":null,"evidence_quote":"Hypersim (plus vKITTI [6]) provides the precise synthetic training data with which the method constructs diverse synthetic priors and trains the conditioned MDE to rectify measurement noise."}],"review_version":1}