{"id":"9ae6159a-c78c-4d40-981e-9bec7c3abeac","arxiv_id":"2603.05711","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Any2Full reformulates depth completion as one-stage scale-prompting of a pretrained monocular depth estimator, yielding domain-general, pattern-agnostic dense metric depth with lower error and higher speed than prior methods.","lead":"Any2Full turns incomplete depth maps into dense metric depth by prompting a frozen monocular depth model with scale cues extracted from the sparse measurements. It is faster and more robust across sensors and scenes than prior two-stage fusion methods, and has already been deployed for warehouse robot grasping.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The central claim is empirical and well-supported by zero-shot tables, efficiency numbers, ablations, and a real robotic deployment. The scale-consistency premise is the load-bearing assumption, exactly as the reader identified, yet the paper already tests it with visualizations, module ablations, and pattern-robustness experiments. No hidden circularity, missing baseline, or formal gap that would force a verdict change appears. Code release further lowers correctness risk. Therefore the ACCEPT / high-confidence verdict stands without adjustment.","tokens_in":23382,"tokens_out":414,"duration_ms":5559,"concrete_test":"Recompute the region-wise scale-factor variance (Supp. A.1 protocol) on the full IBims-1 Range and KITTI-4L subsets for both the frozen DA-L backbone and Any2Full; if the prompted variance remains >5\times lower than the backbone and AbsREL stays within 10% of Tab. 1, the scale-consistency premise holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (that SAPE + multi-level FiLM can render MDE relative depth globally scale-consistent so that a single closed-form LS alignment recovers metric depth) is the correct soft spot, but the paper already supplies direct empirical support for it: Supp. A.1 region-wise scale maps show near-uniform factors after prompting versus the backbone's spatially varying field; Tab. 4 ablations show progressive gains from LE/GP; and zero-shot results hold across Hole/Range/Sparse/Mixed patterns and six domains (Tab. 1). No internal contradiction or missing control that would collapse the one-stage claim is present. The residual risk is only that residual local scale variation could still exist on untested extreme patterns, yet that risk is already quantified by the existing Range and sparsity sweeps and does not undermine the reported averages or the 1.4\times efficiency claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Any2Full, a one-stage depth completion framework that reformulates the task as scale-prompting adaptation of a frozen pretrained monocular depth estimation (MDE) backbone (Depth Anything v2). A lightweight Scale-Aware Prompt Encoder (SAPE) with Local Enrichment (FiLM coupling of sparse scale cues to MDE features) and Global Propagation (MDE-geometry-guided attention) modules converts normalized sparse depth into multi-level scale prompts that modulate the MDE decoder, producing globally scale-consistent relative depth that is recovered to metric depth by a single closed-form least-squares alignment. The method is trained only on synthetic data with random/hole sampling and claims domain-general, pattern-agnostic performance. Extensive zero-shot experiments across six public datasets, six depth patterns (including unseen Range), three backbone sizes, sparsity/range ablations, module ablations, and a real warehouse robotic grasping deployment support superior average AbsREL (32.2% better than OMNI-DC) and 1.4× speedup versus PriorDA with the same backbone.","tokens_in":23612,"tokens_out":810,"duration_ms":20428,"significance":"If the empirical claims hold, Any2Full offers a practical new paradigm for universal depth completion: it inherits MDE geometric priors without two-stage coarse-to-fine overhead or iterative test-time optimization, while remaining robust to sensor-induced sparsity, holes, range limits, and domain shifts. The real-world warehouse deployment (grasp success rising from ~28–39% to 91–100% on black packages) and public code/checkpoints strengthen impact for robotics. Strengths include the clean one-stage formulation, direct scale-consistency evidence (Supp. A.1 region-wise maps), progressive module ablations (Tab. 4), attention visualizations, multi-backbone consistency, and efficiency numbers that make the approach deployable. These elements make the work a solid contribution beyond incremental fusion of existing MDE models.","major_comments":[],"minor_comments":[{"comment":"Abstract and Sec. 4.2 claim a 32.2% average AbsREL improvement over OMNI-DC; Tab. 1 AVG column shows 0.014 → 0.010 (~28.6%). Clarify the exact averaging procedure (per-pattern relative gains, weighted, etc.).","section":null},{"comment":"Fig. 2 caption and diagram contain residual LaTeX artifacts (“Frezze”, “F!”#”, “F+%&%)*” etc.). Clean for camera-ready.","section":null},{"comment":"Sec. 3.2 and Supp. C.2: the multi-size patch embedding and masked attention details are useful; a short pseudocode block or diagram inset would improve reproducibility without expanding the main text.","section":null},{"comment":"Tab. 1 rank computation (Avg Rank 2.3) is not fully specified (how ties or missing entries are handled). A one-sentence note would suffice.","section":null},{"comment":"Fig. 4 and Tab. 3 captions could more explicitly state that Range and extreme sparsity protocols are zero-shot for all methods, reinforcing the pattern-agnostic claim.","section":null},{"comment":"Minor notation: the tilde for normalized relative depth is introduced late; define ˜Ds earlier in Sec. 3.1 for readers skimming equations.","section":null}],"recommendation":"accept","confidential_remarks":"Strong empirical paper with real deployment and code; fits well for a top CV venue. No novelty or citation concerns. The empty major-comments list is intentional—the central one-stage scale-prompting claim is well-supported by the existing controls and does not require further load-bearing fixes."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: they stop doing the usual two-stage relative-to-metric alignment that warps MDE geometry, and instead inject sparse scale cues as multi-level FiLM prompts so the frozen Depth Anything backbone itself produces a globally consistent relative map. One closed-form LS fit then recovers metric depth. That is the whole claim, and the tables support it.\n\nWhat is new is the Scale-Aware Prompt Encoder: Local Enrichment couples sparse ratios into MDE features, then Global Propagation diffuses them with attention whose Q/K come only from MDE geometry (Value from the scale features). Multi-level FiLM follows. Prompting and FiLM are not new, but this particular hierarchical, geometry-guided construction for pattern-agnostic completion is, and it is cleanly engineered. They freeze the backbone, train only the light encoder, release code, and show the expected progressive gains when LE and GP are added (Tab. 4). Supp. A.1 region-wise scale maps are the right control: after prompting the scale field is nearly uniform, which is exactly what the one-stage story needs.\n\nThe empirical package is solid. Zero-shot across six datasets and six patterns (including an unseen Range pattern), three backbone sizes, sparsity/range sweeps, efficiency numbers (1.4\times vs PriorDA, same DA-L), and a real warehouse deployment that lifts black-package grasp success from 28 % to 91.6 %. Average AbsREL beats OMNI-DC by the claimed 32 %. No circular math; the final LS step is just post-processing.\n\nSoft spots are minor and already quantified. Residual local scale variation could still appear on more extreme untested patterns, but the Range and sparsity ablations already stress that axis. Loss weights and layer indices are free parameters, yet they are set uniformly and the ablations are transparent. Citation pattern is fair; they correctly diagnose the two-stage distortion problem.\n\nThis is for anyone building robotic or AR depth pipelines who wants a single feed-forward model that does not need per-sensor fine-tuning. It deserves a serious referee. I would engage with it and expect it to be cited.","headline":"Clean one-stage scale-prompting of Depth Anything that actually beats two-stage MDE baselines on accuracy, speed, and pattern robustness, with real warehouse numbers to back it.","tokens_in":24234,"tokens_out":535,"would_cite":true,"duration_ms":5405,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Any2Full recovers dense metric depth from any incomplete sensor map by scale-prompting a frozen monocular depth model in one forward pass.","keywords":["depth completion","monocular depth estimation","scale prompting","domain generalization","pattern-agnostic","FiLM modulation","robotic perception"],"falsifier":"On a held-out scene whose sparse measurements leave large unmeasured regions, compute the variance of local scale factors needed to align the prompted relative depth to ground-truth metric depth; if that variance remains large, the closed-form global fit fails and the one-stage claim collapses.","tokens_in":24302,"feed_emoji":"📏","tokens_out":614,"duration_ms":9203,"temperature":0.7,"pith_summary":"Commodity depth sensors return sparse or gappy maps that break robotic perception. Traditional completion networks overfit both the RGB domain and the exact hole or sparsity pattern they saw in training, while two-stage methods that bolt a monocular depth estimator onto an explicit alignment step add compute and inject structured artifacts. Any2Full instead freezes a strong monocular depth backbone and learns only a lightweight Scale-Aware Prompt Encoder that turns the sparse metric measurements into multi-level scale prompts. These prompts modulate the backbone so its relative depth becomes globally scale-consistent and can be converted to metric depth by a single closed-form fit. The result is a domain-general, pattern-agnostic completer that is both more accurate and faster than prior approaches, already running on warehouse robots that must grasp black packages whose depth is almost entirely missing.","feed_headline":"One prompt turns any sparse depth into dense metric maps","feed_subtitle":"A frozen monocular model plus lightweight scale cues beats multi-stage completers on accuracy and speed","key_machinery":"Scale-Aware Prompt Encoder (SAPE): a hierarchical module that first anchors local scale ratios from sparse depth into monocular features via FiLM, then propagates those cues across the scene with geometry-guided attention, finally injecting the resulting multi-level prompts into the monocular decoder so the relative depth becomes scale-consistent.","core_discovery":"Depth completion can be reformulated as a one-stage scale-prompting adaptation of a pretrained monocular depth estimator: sparse metric cues are distilled into unified scale prompts that force the model’s relative output to be globally consistent (alignable by one scale and bias) while leaving its geometric priors intact, thereby achieving domain-general and pattern-agnostic dense metric recovery without intermediate coarse maps or iterative refinement.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Scale prompts turn sparse depth into dense metric maps","One-stage scale prompts adapt Depth Anything for full completion","Sparse cues become scale prompts for pattern-agnostic metric depth","Frozen MDE plus scale prompts yields domain-general dense depth","Any-to-full: prompting monocular priors into metric completion"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"After simple global normalization that keeps only inter-point scale ratios, the prompt encoder can turn those sparse ratios into multi-level FiLM signals that make the monocular output truly globally consistent rather than still locally warped.","fun_headline_variants_meta":{"raw":{"variants":["Scale prompts turn sparse depth into dense metric maps","One-stage scale prompts adapt Depth Anything for full completion","Sparse cues become scale prompts for pattern-agnostic metric depth","Frozen MDE plus scale prompts yields domain-general dense depth","Any-to-full: prompting monocular priors into metric completion"]},"model":"grok-4.5","effort":"low","cost_usd":0.005362,"raw_usage":{"total_tokens":1491,"prompt_tokens":800,"num_sources_used":0,"completion_tokens":66,"cost_in_usd_ticks":53620000,"prompt_tokens_details":{"text_tokens":800,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":625,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":800,"tokens_out":66,"duration_ms":5895,"temperature":1.0,"reasoning_tokens":625,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T14:16:58.820229+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out scene whose sparse measurements leave large unmeasured regions, compute the variance of local scale factors needed to align the prompted relative depth to ground-truth metric depth; if that variance remains large, the closed-form global fit fails and the one-stage claim collapses.","supporting_citations":[],"review_version":1}