{"id":"87185514-c5d6-49e5-8990-6dcb5c3d2a38","arxiv_id":"2501.01121","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A lightweight refiner with a coarse-to-fine denoising module, noise-based pretraining, and a scale-shift invariant gradient-matching loss achieves state-of-the-art high-resolution metric depth with up to 10x faster inference.","lead":"PatchRefiner V2 makes high-resolution depth estimation much faster and lighter by replacing a heavy neural network with a small encoder in the refinement stage. It adds a denoising module and a new training loss to preserve accuracy, and reports state-of-the-art results on the UnrealStereo4K benchmark with far fewer parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA and speed claims rest on an unverifiable comparison with reimplemented baselines, and Eq. (6) defines a loss that is identically zero as written; without code or a corrected equation, the headline result cannot be independently checked.","rationale":"The reader's conditional verdict is appropriate, but I would locate the load-bearing risk differently. The paper's internal ablations are consistent across Tables 4, 6, and 7, and the synthetic SOTA claim does not depend on the SSIGM loss, so the coarse-branch reliability assumption is not the primary threat to the headline. The deeper problem is verifiability and baseline fairness: the strongest quantitative comparison is made against reimplemented baselines whose pretraining was deliberately weakened, and the only published formula for a component claimed to drive real-domain gains is internally inconsistent. The 0.9% RMSE margin over the original PatchRefiner is small enough that evaluation details could matter, and the absence of code and checkpoints prevents an independent check. For these reasons I would keep the reader's CONDITIONAL verdict rather than accepting or rejecting: the architecture and ablations are plausible, but the central SOTA/speed claim should be conditional on artifact release and a corrected, verifiable comparison.","tokens_in":16320,"tokens_out":6658,"duration_ms":69213,"concrete_test":"Ask the authors to release code and trained checkpoints, then rerun Table 1 on the same GPU using the original unmodified PatchRefiner and PatchFusion checkpoints (without removing the MiDaS pretraining stage) and also under their aligned protocol. If PRV2C's RMSE advantage over original PatchRefiner shrinks below noise or the 0.62s timing cannot be reproduced, the SOTA and speed claims should be softened. In the same release, the actual SSIGM implementation should show the gradient computed on d^*_i - \\hat d_i, not on a zero residual.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PRV2C achieves a new state of the art on UnrealStereo4K (RMSE 0.884, 0.62s, 2.3x faster than PatchRefiner) depends on a comparison whose conditions are neither fully specified nor publicly checkable. Table 1 compares PRV2 against 'aligned' versions of PatchFusion and PatchRefiner obtained by removing a non-public MiDaS pretraining stage from the refiner/fine branch, while leaving the coarse branch unchanged. The original PatchRefiner reports RMSE 0.892, so PRV2C's 0.884 is only a 0.9% improvement over the original, and most of the apparent SOTA margin comes from comparing against the weakened aligned PR (0.941). This is load-bearing because a small change in evaluation protocol, training randomness, or the alignment procedure could erase the claimed advantage. Separately, Eq. (6) defines the proposed SSIGM loss with R_i = \\hat d_i - \\hat d^*_i, but Eq. (4) sets \\hat d^*_i = \\hat d_i, so R_i is identically zero and the loss is zero as written. The real-domain boundary improvements attributed to SSIGM therefore rely on an implementation detail that the paper does not correctly specify. With no released code or checkpoints and a modified baseline protocol, the reported speed-accuracy tradeoff cannot be independently reproduced or attributed to the proposed components.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PatchRefiner V2 (PRV2), a tile-based high-resolution metric depth estimation framework that replaces the heavy refiner branch of PatchRefiner with lightweight encoders. To compensate for the resulting loss of depth-aligned features, the authors propose a Coarse-to-Fine (C2F) module containing Guided Denoising Units (GDUs), a Noisy Pretraining (NP) strategy, and a Scale-and-Shift Invariant Gradient Matching (SSIGM) loss for synthetic-to-real transfer. They evaluate on UnrealStereo4K and on Cityscapes, ScanNet++, and KITTI, and claim that PRV2C achieves a new state of the art on UnrealStereo4K with RMSE 0.884 while being 2.3x faster than PatchRefiner, with improved boundary delineation on real-domain datasets.","tokens_in":16709,"tokens_out":5096,"duration_ms":50113,"significance":"If the claims are validated, the paper makes a useful practical contribution: it shows that a lightweight refiner with a learned denoising module can recover much of the quality of a heavy refiner while drastically reducing parameter count and inference time. The ablations in Tables 4, 6, and 7 consistently show monotonic improvements from end-to-end training, C2F, and NP across three model sizes, which is a genuine strength. The evaluation uses external datasets and external base models, and the use of the previous PatchRefiner model as the pseudo-label teacher is standard self-training rather than circular reasoning. However, two load-bearing issues currently prevent the headline claims from being accepted: the SSIGM loss as written in Eq. (6) is identically zero, and the state-of-the-art comparison in Table 1 relies on reimplemented, 'aligned' baselines whose details are not fully specified and whose margin over the original published PatchRefiner is very small.","major_comments":[{"comment":"The SSIGM loss is defined with R_i = \\hat d_i - \\hat d^*_i, but Eq. (4) states \\hat d^*_i = \\hat d_i. Therefore R_i is identically zero and L_ssigm in Eq. (6) is identically zero as written. This makes the SSIGM contribution vacuous on paper and means the real-domain boundary gains reported in Tables 3 and 5 cannot currently be attributed to the proposed loss. Please correct the definition (presumably R_i should be d^*_i - \\hat d^*_i) and confirm that the experiments used the corrected formulation.","section":"§3.3, Eq. (6)"},{"comment":"The headline claim that PRV2C sets a new state of the art rests on the 'aligned' baselines PR† and PF†, which are obtained by removing a non-public MiDaS pretraining stage from the refiner or fine branch. The original PatchRefiner row reports RMSE 0.892, while PRV2C reports 0.884, a difference of only 0.008 (about 0.9%); the gap to the aligned PR† is 0.057. Because the alignment procedure is not described in enough detail, no trained models or checkpoints are released, and no multiple-seed variance is reported, the claimed margin over the actual published state of the art is not verifiable and may be within run-to-run variation. Please report results against the original published PatchRefiner and PatchFusion models, provide error bars or multiple runs, and release code/checkpoints.","section":"§4.3, Table 1"},{"comment":"The caption states that #param and T denote the additional parameters and the inference time of the fine or refiner branch only. The coarse branch is shared across compared methods, so the end-to-end inference speedup is smaller than the reported 2.3x. Since the abstract claims that PRV2 outperforms previous methods in both accuracy and speed, the paper should report total end-to-end inference time (coarse plus refiner) or clearly qualify the speed comparison as refiner-branch-only.","section":"§4.3, Table 1 caption"},{"comment":"The Noisy Pretraining strategy replaces input coarse features with random normal noise, and the ablation Table 4 attributes a substantial RMSE improvement to this component. However, the paper does not specify the noise scale, whether the noise is resampled at every optimization step, or how the random features are matched to the resolution of each C2F layer. Without these details, the NP component cannot be reproduced from the manuscript.","section":"§3.2.3"}],"minor_comments":[{"comment":"There are several typos and naming inconsistencies: 'CityScape' should be 'Cityscapes'; 'Midas' should be 'MiDaS'; 'Convnext' should be 'ConvNeXt'; and in §3.3 'predicted depthdi' lacks a space between 'depth' and 'd_i'.","section":"Throughout"},{"comment":"The column header 'Data pl gen' is unclear; please spell out that 'pl' denotes pseudo-labels and 'gen' denotes their online or offline generation.","section":"Table 3"},{"comment":"The notation 'GMSSI' is defined only indirectly; please state explicitly that GMSSI means gradient maps are computed first and then aligned with LSE, in contrast to SSIGM, so that the order-of-operations ablation is self-contained.","section":"Table 5 caption"},{"comment":"The KITTI setup cites Garg et al. for cropping to 352x1216, but the usual Eigen split evaluation uses a specific crop; please clarify whether the reported protocol matches the Eigen split exactly or deviates from it.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea is interesting, but the two main technical claims need substantiation: the corrected SSIGM loss must be stated and verified, and the SOTA comparison must be placed on firmer ground (preferably against the original published checkpoints, with code release). The authors should also clarify the speed metric. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PRV2 is a credible engineering follow-up to PatchRefiner. The core idea—replace the heavy refiner with a MobileNet/EfficientNet/ConvNext encoder and add a coarse-to-fine module with gated denoising plus noisy pretraining—is sensible, and the ablations show each component buys a monotonic RMSE improvement. The real-domain boundary F1 gains on CityScapes (+17–21% over the prior SSI/ranking losses) are the most convincing result in the paper, especially since they replicate on both PR and PRV2E.\n\nThe soft spots are real but not fatal. Equation (6) defines SSIGM with R_i = \\hat d_i - \\hat d^*_i, while Eq. (4) sets \\hat d^*_i = \\hat d_i, so the loss is identically zero as written. That has to be a typo for d^*_i - \\hat d^*_i, but it is a load-bearing typo: the paper's main new loss is not correctly specified. The abstract also overclaims: PRV2M and PRV2E are worse than the original PatchRefiner on RMSE (1.003 and 0.948 vs 0.892), and only PRV2C beats it, by a slim margin (0.884 vs 0.892). The 'aligned' baselines in Table 1 exaggerate the gap by removing the MiDaS pretraining from PR/PF; that is a defensible protocol choice, but the reader should see the original numbers too. Finally, no code or checkpoints are provided, so none of this can be independently checked.\n\nThe coarse-branch reliability assumption is a genuine limitation, and the paper could be clearer about failure cases when D_c is poor. Still, the engineering contribution is solid, the ablations are honest, and the boundary results are strong. A serious referee should engage with this after the authors fix Eq. (6), release code, and tone down the abstract. I would not cite it in its current form, but I'd keep an eye on the revision.","headline":"Solid engineering follow-up to PatchRefiner, but Eq. (6) is identically zero as written and the SOTA claim is overstated; worth reviewing after fixes.","tokens_in":17224,"tokens_out":3593,"would_cite":false,"duration_ms":30091,"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":"A lightweight refiner guided by coarse depth features can beat heavy tile-based refinement on 4K depth accuracy while cutting inference time.","keywords":["monocular depth estimation","high-resolution depth","tile-based refinement","coarse-to-fine module","Guided Denoising Unit","Noisy Pretraining","scale-and-shift invariant gradient matching","synthetic-to-real transfer"],"falsifier":"Run PRV2 on a real-domain collection where the coarse base model produces systematically biased depth (for example, night-driving or rain images outside its training distribution) and test whether the refiner removes the bias; the paper's premise predicts it will not, because the refiner only adds local residuals to the coarse depth map.","tokens_in":16127,"feed_emoji":"📏","tokens_out":8911,"duration_ms":71207,"temperature":0.7,"pith_summary":"PatchRefiner V2 asks whether the expensive second pass in tile-based high-resolution depth estimation is necessary. The paper's answer is that a lightweight encoder can replace the heavy depth foundation model in the refiner branch, provided the lost depth-aligned features are restored by two new mechanisms: a coarse-to-fine module that uses global coarse-depth features to denoise local patch features, and a pretraining stage that feeds the refiner random noise instead of coarse features. On the synthetic 4K benchmark UnrealStereo4K, the largest variant PRV2C achieves an RMSE of 0.884, a new state-of-the-art, while running 2.3x faster than the previous PatchRefiner; the smallest variant is 9.2x smaller and 10.7x faster than PatchFusion. A third contribution, a scale-and-shift invariant gradient-matching loss for synthetic-to-real transfer, improves boundary F1 scores on Cityscapes by 17.2% for PatchRefiner and 20.9% for PRV2E. If these results hold, high-resolution metric depth estimation no longer needs to trade accuracy for speed.","feed_headline":"4K depth estimation gets 2.3x faster and more accurate","feed_subtitle":"A lightweight refiner with guided denoising cuts parameters and sharpens depth edges on real scenes.","key_machinery":"The central mechanism is the Coarse-to-Fine (C2F) module, a bottom-up stack of layers each containing a Guided Denoising Unit (GDU) and a residual convolutional unit. The GDU concatenates coarse features with fine refiner features, passes them through a convolutional block and sigmoid to produce a 0-to-1 weight map, and multiplies that map into the fine features, denoising them before the Fine-to-Coarse (F2C) module injects fine detail into the coarse depth map. Around this sit two supporting mechanisms: Noisy Pretraining, which feeds random normal noise as coarse-guidance input so the refiner branch learns to extract depth-relevant features on its own, and the Scale-and-Shift Invariant Gradient Matching (SSIGM) loss, which aligns predicted and pseudo-label depths by least-squares scale/shift and then penalizes gradient differences so high-frequency boundaries are learned directly.","core_discovery":"On the paper's own terms, the discovery is that the refinement stage of tile-based depth estimation can be made fast and small without sacrificing accuracy, because what the lightweight refiner lacks is not capacity but an aligned feature representation. The C2F module supplies that alignment by letting coarse features act as a gating signal: a Guided Denoising Unit computes a sigmoid weight map from concatenated coarse and fine features and multiplies it into the fine shortcut features, suppressing noise while preserving high-frequency structure. Noisy Pretraining then pretrains the whole refiner branch with random coarse features, forcing it to learn depth-relevant features without relying on the coarse branch. Finally, SSIGM replaces the scale-and-shift invariant MAE with a gradient-matching term after least-squares scale/shift alignment, so pseudo-label supervision directly targets high-frequency detail. The combination yields state-of-the-art RMSE 0.884 on UnrealStereo4K, larger boundary F1 gains on Cityscapes, and consistent improvements when the base model is either ZoeDepth or Depth Anything V2.","pith_inferences":["The Noisy Pretraining recipe may transfer to other dual-branch dense prediction tasks where a global branch guides local processing; its only requirement is that guidance features can be replaced by random noise during pretraining.","The SSIGM design suggests a general ordering principle for gradient-based domain transfer: align global scale and shift first, then match gradients; the paper's ablation shows the reverse order (GMSSI) is worse on boundary metrics.","A testable implication of the coarse-branch assumption is that PRV2's advantage should shrink when the coarse base model is unreliable in the target domain, since the refiner only adds residuals to $D_c$ and cannot repair global bias.","System-level latency may not scale exactly with refiner parameter count, because patch cropping, ROI extraction, and consistency-aware inference add fixed overheads beyond the refiner forward pass."],"forward_implications":["PRV2C sets a new state-of-the-art RMSE of 0.884 on UnrealStereo4K while running 2.3x faster than PatchRefiner.","PRV2M shows that a 47.0M-parameter refiner can improve RMSE by 22.2% over the coarse baseline while being 9.2x smaller and 10.7x faster than PatchFusion.","The SSIGM loss improves boundary F1 by 17.2% on PatchRefiner and 20.9% on PRV2E on Cityscapes without hurting scale accuracy.","The framework is base-model agnostic: it improves both ZoeDepth and Depth Anything V2 coarse branches on UnrealStereo4K.","End-to-end training of the full framework becomes feasible, replacing the stage-wise training required by earlier tile-based methods."],"supporting_citations":[{"why":"PatchRefiner V1 is the baseline architecture that PRV2 replaces and extends, providing the tile-based coarse/fine refinement and DSD training setup.","marker":"[36]"},{"why":"PatchFusion supplies the end-to-end tile-based framework and consistency-aware inference used for comparison and adopted in PRV2.","marker":"[34]"},{"why":"ZoeDepth is the coarse base model used for the main UnrealStereo4K comparisons and for pseudo-label generation.","marker":"[4]"},{"why":"UnrealStereo4K (SMD-Net) provides the synthetic 4K dataset and stereo ground truth on which the main accuracy claims are measured.","marker":"[65]"},{"why":"The scale-and-shift invariant loss and least-squares alignment procedure are the starting point that SSIGM modifies.","marker":"[54]"},{"why":"MegaDepth supplies the gradient-matching loss that SSIGM combines with scale-and-shift alignment.","marker":"[32]"},{"why":"MobileNet is the lightweight encoder used in the smallest PRV2 variant, PRV2M.","marker":"[51]"},{"why":"EfficientNet is the lightweight encoder used in the middle PRV2E variant.","marker":"[64]"},{"why":"Depth Anything V2 serves as the alternative coarse base model that demonstrates the framework's versatility in Table 2.","marker":"[72]"},{"why":"ImageNet pretraining for the lightweight encoder is the starting point whose insufficiency motivates Noisy Pretraining.","marker":"[13]"}],"fun_headline_variants":["PatchRefiner V2: 2.3x faster depth with fewer parameters","Lightweight refiner with guided denoising speeds up 4K depth","Noisy pretraining yields accurate, lightweight depth estimation","Small refiner, sharp edges: 4K depth gets fast and light"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The coarse branch already provides a reliable base depth estimate, so the refiner only needs to add high-frequency detail; if the coarse depth is badly wrong in a target domain, the lightweight refiner cannot repair it.","fun_headline_variants_meta":{"raw":{"variants":["PatchRefiner V2: 2.3x faster depth with fewer parameters","Lightweight refiner with guided denoising speeds up 4K depth","Noisy pretraining yields accurate, lightweight depth estimation","Small refiner, sharp edges: 4K depth gets fast and light"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001442,"raw_usage":{"total_tokens":5816,"prompt_tokens":959,"completion_tokens":4857,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":4778}},"tokens_in":575,"tokens_out":4857,"duration_ms":31454,"temperature":1.0,"reasoning_tokens":4778,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:34:27.878008+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PRV2 on a real-domain collection where the coarse base model produces systematically biased depth (for example, night-driving or rain images outside its training distribution) and test whether the refiner removes the bias; the paper's premise predicts it will not, because the refiner only adds local residuals to the coarse depth map.","supporting_citations":[{"cited_title":"Smd-nets: Stereo mixture density networks","cited_arxiv_id":null,"evidence_quote":"UnrealStereo4K (SMD-Net) provides the synthetic 4K dataset and stereo ground truth on which the main accuracy claims are measured."},{"cited_title":"Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer","cited_arxiv_id":null,"evidence_quote":"The scale-and-shift invariant loss and least-squares alignment procedure are the starting point that SSIGM modifies."},{"cited_title":"Megadepth: Learning single- view depth prediction from internet photos","cited_arxiv_id":null,"evidence_quote":"MegaDepth supplies the gradient-matching loss that SSIGM combines with scale-and-shift alignment."},{"cited_title":"Imagenet: A large-scale hierarchical image database","cited_arxiv_id":null,"evidence_quote":"ImageNet pretraining for the lightweight encoder is the starting point whose insufficiency motivates Noisy Pretraining."}],"review_version":1}