{"id":"38e37643-7d07-4f98-8e06-6589631b8773","arxiv_id":"2506.17110","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"One-shot scale-rotation-shift alignment of monocular depth model output, guided by sparse ground-truth depth points, yields metric depth accurate enough for robot grasping.","lead":"This paper introduces a one-time calibration that turns a single RGB camera image into metric depth for robot grasping, by aligning the output of a monocular depth model to a handful of true depth points. The method reports high success rates on tabletop grasping and bin-picking, including transparent objects that confuse normal depth sensors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"One-shot transfer is the load-bearing assumption: min-max normalization (Eq. 12) removes only global scale/shift, so a map fit on static background may not extrapolate to object depths; Tab. I should be re-run with a disjoint calibration/evaluation split.","rationale":"The paper has independent support: real robot trials (Tab. V), a code link, and a well-posed linear alignment formulation. The most insecure premise is the transfer of a fixed alignment across changing scene content, which the authors themselves flag in Sec. IV-C and Fig. 4. I agree with the reader's weakest_assumption. The proposed concrete split test would settle it: if object-mask metrics survive a disjoint-scene one-shot protocol, the main claim stands; if not, the paper is better described as a within-scene calibration method. Secondary issues, such as the Eq. (12) normalization typo and lack of error bars, are real but do not change the conditional verdict; they reinforce the need for a careful re-evaluation rather than rejection. Thus the reader's CONDITIONAL verdict remains appropriate.","tokens_in":12663,"tokens_out":7458,"duration_ms":85509,"concrete_test":"Re-run the Pose 1-4 evaluation under a strict one-shot protocol: use one calibration image whose object configuration is disjoint from all evaluated images, sample the 100 calibration points only from the static table/tote mask, and report object-mask delta_1.10 and MAE for the held-out scenes. As a sharper check, fit Eq. (6) twice on the same held-out image, once from static-background points and once from object points, and compare the predicted object depth; if the two fits disagree by more than the claimed 0.03 MAE margin, the transfer assumption fails. Also recompute Table II using the standard min-max formula (z_p - z_min)/(z_max - z_min) and compare with Eq. (12); if any entry shifts by more than 10%, the printed normalization formula is materially affecting the results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, that SSRA turns one calibration image into metric depth for all later images in a fixed setup, requires that after min-max normalization the MDEM output z_p is linked to metric depth z_c by the same function F(·|Theta_hat) in Eq. (6) across trials. Section IV-C concedes that raw predicted depth fluctuates with scene content (Fig. 4) and offers normalization as the fix, but the global min/max rescaling in Eq. (12) compensates only for a single global scale and shift; it cannot remove spatially varying, content-dependent error. The calibration points are described as sampled from the static portion of the scene (e.g., the table), while the metrics in Tab. I and Tab. IV are computed on object masks. If the MDEM's error behavior on objects is not the same linear function of z_p, u*z_p, v*z_p as on the static background, the fitted map transfers nothing about the objects. Moreover, the evaluation protocol does not clearly state that the calibration RGB/D image is a different scene from the 100 evaluation images per pose; if calibration is drawn from the same object configurations, Tab. I is a within-scene fit, not a one-shot generalization. The headline quantitative claim \"delta_1.10 > 0.85 and MAE < 0.03 in all settings\" is therefore sensitive to the exact scene-split protocol. Secondary but relevant: Eq. (12) as printed adds z_min back, so it does not map to a fixed [0,1] range; this typo must be corrected before trusting the normalization ablation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MOMA, a framework for recovering metric depth from a single RGB image by aligning the output of a monocular depth estimation model (MDEM) to sparse ground-truth depth points through a one-shot scale-shift-rotation alignment, instantiated as SSRA. The alignment parameters are fit on one calibration image and then applied to subsequent images in a fixed camera setup. The authors evaluate SSRA against global scale-shift alignment (GSSA), locally weighted linear regression (LWLR), and raw DAM output on opaque and transparent objects under multiple camera poses, report ablation studies on the number of calibration points, normalization, and MDEM choice, and present real-robot two-finger grasping and suction-based bin-picking experiments.","tokens_in":13004,"tokens_out":4271,"duration_ms":46032,"significance":"If the one-shot transfer claim holds, the contribution is practically valuable: it offers a simple, low-cost route to metric depth for fixed-camera robot manipulation without a continuously operating depth sensor. The method is straightforward, the paper includes real-robot experiments, and the authors make code and data available. The positive aspects are the clear formulation of the alignment problem, the explicit treatment of MDEM output fluctuation via normalization, and the demonstration that even a simple calibrated affine model can substantially improve raw MDEM depth. However, the evaluation protocol currently leaves the central transfer claim under-supported, and there are several concrete errors in the reported formulas and results that need correction before the claims can be accepted.","major_comments":[{"comment":"The one-shot transfer protocol is underspecified. Section IV-C states that calibration relies on a static portion of the scene (e.g., the table), while Section V-A says calibration uses \"an RGB/D image\" with \"100 samples from the depth image\" but does not state whether that image is disjoint from the 100 evaluation images or whether the samples lie only on the static background. Since the metrics in Table I are computed on object masks, the current description cannot rule out that the reported numbers partly reflect within-scene fitting rather than transfer to new scenes. Please specify the exact scene/object split and rerun Table I with a calibration image and calibration point set that are disjoint from the evaluation images, and report results for calibration points restricted to the static background versus points sampled over the full scene, to substantiate the one-shot generalization claim.","section":"§IV-C, §V-A, Table I"},{"comment":"Equation (12) as printed is not a min-max normalization to a fixed range: it reads \\hat z_p = (z_p - z_min)/(z_max - z_min) + z_min, which maps the data to [z_min, z_min+1] rather than [0,1]. This contradicts the stated purpose of stabilizing MDEM output fluctuations and the interpretation of the ablation in Table II. Please correct the equation (presumably removing the final + z_min) and verify that the implementation matches the corrected formula.","section":"§IV-C, Eq. (12)"},{"comment":"The text states that for the tote scenario \"the δ1.05 score is above 0.85 and the mean absolute error (MAE) for the tote scenario is 0.0242,\" but Table IV reports SSRA in the tote scenario with δ1.05 = 0.5708, δ1.10 = 0.8547, and MAE = 0.0242. The sentence appears to conflate δ1.05 and δ1.10. Please correct the text to match the table.","section":"§V-E, Table IV"},{"comment":"The 50-trial success rates in Table V are reported without confidence intervals or significance tests. For n=50, the standard error of a proportion is roughly 5-7 percentage points, so differences such as 82% vs 72% for opaque two-finger grasping and 86% vs 86% for suction grasping are not clearly distinguishable from sampling noise. Please report binomial confidence intervals or repeated-run statistics, and temper the claims that SSRA \"beats\" the baselines accordingly.","section":"§V-F, Table V"}],"minor_comments":[{"comment":"The sentence \"For GSSA, b=100 is used in the evaluation\" appears to be a typo: the bandwidth b is a parameter of LWLR in Eq. (3), not of GSSA. Please correct.","section":"§V-A"},{"comment":"In Eq. (3), the notation Z_p = [z_p,1]^T is not dimensionally consistent with the subsequent weighted least-squares expression; Z_p should be an n×2 design matrix with a column of ones. Please fix the notation.","section":"§IV-B, Eq. (3)"},{"comment":"The abstract says MOMA works \"without additional data collection or model retraining,\" but Section V-E describes fine-tuning DAM on roughly 500K transparent-object images. Please clarify that the one-shot calibration applies to the target setup, while the optional transparent-object fine-tuning is a separate training stage.","section":"§I, Abstract"},{"comment":"The phrase \"using 100 samples from the depth image\" is ambiguous about whether the samples are random, uniformly spaced, or selected from a particular scene region; please specify the sampling procedure.","section":"§V-A"}],"recommendation":"major_revision","confidential_remarks":"The core idea is simple and plausible, and the paper addresses a real practical problem in robot manipulation. However, I am not fully convinced by the one-shot transfer evaluation as currently described: the calibration/evaluation split is not stated precisely, and the transfer from static-background calibration points to object-mask metrics is exactly the load-bearing assumption. The errors in Eq. (12) and in the Section V-E result summary are also concerning and should be fixed before the paper can be considered. I do not see this as a reject, because the method and experiments appear salvageable, but the re-analysis needs to be done carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, it is a practical engineering result: the authors show that a simple linear alignment of a monocular depth model's output, fit on one RGB-D frame, gives metric depth accurate enough for real grasping, including transparent objects, with success rates in the 70-86% range. Second, the method is much closer to existing scale-shift alignment than the 'rotation' framing suggests, and the paper has a few errors and protocol gaps that need fixing.\n\nThe core idea is well-posed: after normalizing the MDEM output, they fit a per-pixel affine map from {u*z_p, v*z_p, z_p} to metric depth using sparse ground truth. That is effectively a four-parameter linear least-squares fit. The rotation angles and pseudo-intrinsics make a nice story, but algebraically they just reparameterize the same linear family. Substituting the pinhole model into Eq. (6)-(8) gives exactly a*u*z_p + b*v*z_p + c*z_p + T3. So the novelty is in the application (one-shot, fixed camera, grasping), not the model class.\n\nWhat is genuinely good: the evaluation is honest in scope. Four camera poses, ablations on sample count and normalization, and real robot trials with 50 trials per scenario. The method beats GSSA and LWLR on most metrics, and the advantage is consistent. The transparent-object fine-tuning of DAM with LoRA is a sensible contribution.\n\nSoft spots, in order of softness. Eq. (12) is miswritten: adding z_min at the end means it is not mapping to [0,1], so either the formula or the claim about min-max normalization needs correcting. Second, the 50-trial success rates have no confidence intervals; with n=50, the gap between 82% and 72% is within noise. Third, the calibration/evaluation protocol in Tab. I is under-specified. The paper says one RGB/D image is used to compute alignment parameters, applied to 'the rest of the evaluations,' but it never states that the calibration image is disjoint from the evaluation images in scene content. If the calibration image comes from the same object placements, the one-shot generalization claim is weaker. The stress-test concern about background versus object masks is also legitimate: if the 100 depth samples are mostly from the static table, the fitted map may not transfer to object depths. The good results suggest it does for these objects and poses, but the paper should demonstrate this explicitly.\n\nOverall, a solid engineering paper with a clear mechanism and useful results. The flaws are fixable and not load-bearing. I would send it to peer review: a good referee will ask for the corrected formula, confidence intervals, and a clearly disjoint calibration/evaluation split. Not a landmark, but a useful data point for RGB-only manipulation.","headline":"A practical one-shot linear alignment for RGB-only grasping that works in real robot trials, but it is closer to a reparameterized scale-shift fit than the 'rotation' framing suggests, and the evaluation protocol needs tightening before the one-shot claim is fully supported.","tokens_in":13567,"tokens_out":3431,"would_cite":true,"duration_ms":34764,"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 a robot can recover metric depth from a single RGB image by applying a one-shot scale–rotation–shift correction to a monocular depth model's output, calibrated with sparse ground-truth depth points, and demonstrates…","keywords":["monocular depth estimation","metric depth alignment","one-shot calibration","scale-rotation-shift alignment","robotic grasping","transparent objects","affine-invariant depth"],"falsifier":"Measure the aligned depth error on object masks across many scene variations using a single calibration, and compare with re-fitting the transform per frame; a large gap would mean the one-shot transform does not transfer to novel objects.","tokens_in":12429,"feed_emoji":"🤖","tokens_out":6920,"duration_ms":66021,"temperature":0.7,"pith_summary":"The paper claims that a robot can recover metric (meter-accurate) depth from a single RGB image by fitting one scale–rotation–shift correction to the output of a monocular depth estimation model, using a sparse set of ground-truth depth points captured once during calibration. The concrete realization, called SSRA, optimizes a seven-parameter transform on min-max normalized predicted depth. On tabletop grasping and bin-picking with a real robot arm, the method reaches 82–86% success on opaque objects and 70–78% on transparent ones, while the same pipeline running on RGB-D fails on transparent objects. This matters because it offers a depth-sensor-free route to accurate depth for fixed-camera manipulation, avoiding costly sensors and their failure modes. The paper positions SSRA as an initial step rather than a complete replacement for RGB-D systems.","feed_headline":"One-shot alignment turns a single RGB image into metric depth","feed_subtitle":"One calibration with a few dozen depth points lets a robot grasp opaque and transparent objects with no depth sensor.","key_machinery":"The load-bearing object is the scale–rotation–shift alignment map $F(u,v,z_p|\\Theta)$, the depth component of a Euclidean transform applied to the point cloud unprojected from the depth model's depth map under a learned pseudo-intrinsic. Concretely, $F = s(-x_p \\sin\\phi + y_p \\sin\\theta\\cos\\phi + z_p \\cos\\theta\\cos\\phi) + T_3$, with $x_p,y_p$ obtained from pixel coordinates and pseudo-intrinsics. Solving the optimization over $\\Theta$ fits the transform to sparse ground-truth depth points once; the same $\\hat\\Theta$ then converts every later depth-model output into metric depth. Min-max normalization of the predicted depth precedes the fit, because the raw depth range fluctuates with scene content even when camera and background are fixed.","core_discovery":"The central discovery is that the affine-invariant depth produced by a modern monocular depth estimator can be promoted to metric depth by a global transform that accounts for scale, rotation of the predicted point cloud, and shift—not just scale and shift. After min-max normalizing the raw depth map, the method solves a nonlinear least-squares problem over $\\Theta=[s,\\theta,\\phi,T_3,c_x^p,c_y^p,f_p]^\\top$ to minimize $\\sum_k \\|z_{c,k} - F(u_k,v_k,z_{p,k}|\\Theta)\\|^2$, where $F$ is the depth component of an $s R X_p + T$ rigid alignment under a pinhole model. With one calibration image and roughly tens to hundreds of ground-truth depth samples, the fitted transform stays valid across later images, giving $\\delta_{1.10}>0.85$ and MAE below 0.03 meters on object masks. The method also extends to transparent objects after fine-tuning the underlying depth model on transparent-object datasets.","pith_inferences":["A natural stress test is to sample calibration points from objects rather than the fixed background; if the fitted transform changes sharply, the one-shot transfer is relying on the background being a good proxy for object depth errors.","Extending the one-shot fit to a few calibration scenes should reduce sensitivity to scene-dependent depth fluctuations; the paper suggests this direction but does not quantify how many scenes are needed.","Because the fitted parameters include pseudo-intrinsics, the rotation terms may be absorbing a systematic bias in the depth model's virtual camera rather than a physical scene rotation; isolating the dominant parameter could lead to cheaper calibration."],"forward_implications":["Fixed-camera industrial settings such as bin picking can use a single RGB camera and a one-time calibration to obtain depth accurate enough for grasping.","The alignment fits any monocular depth model, so swapping the underlying model only changes the fitted parameters, not the pipeline.","Fine-tuning the depth model on transparent-object data lets the one-shot calibration extend to transparent objects, which structured-light and time-of-flight sensors fail on.","Calibration takes a few seconds and runtime alignment takes under 10 milliseconds per frame, so the approach is compatible with real-time control loops.","Accuracy saturates at a few hundred ground-truth samples, making calibration data collection inexpensive."],"supporting_citations":[{"why":"Supplies the base monocular depth estimation model whose output is aligned, and is evaluated as the raw metric-depth baseline.","marker":"[7]"},{"why":"Cited as part of the global scale-shift alignment baseline and its least-squares fitting.","marker":"[6]"},{"why":"Supplies the median-based normalization variant and is cited for global scale-shift fitting.","marker":"[27]"},{"why":"Supplies the locally weighted linear regression alignment baseline.","marker":"[28]"},{"why":"Provides transparent-object depth-completion training data used to fine-tune the depth model.","marker":"[20]"},{"why":"Provides transparent-object depth-completion training data used to fine-tune the depth model.","marker":"[21]"},{"why":"Supplies the scale-invariant loss used when fine-tuning the depth model on transparent objects.","marker":"[33]"},{"why":"Supplies the min-max normalization that precedes the alignment fit.","marker":"[29]"}],"fun_headline_variants":["One-shot depth alignment gives robots metric depth from RGB","MOMA: single RGB image to metric depth in one shot","Scale-rotation-shift alignment unlocks metric depth from one RGB","No depth sensor: one-shot RGB metric depth for grasping","MOMA converts a single RGB image into grasp-ready metric depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a transform fitted once to sparse points on the fixed background still corrects the depth model's output on objects that appear later, even though the model's raw depth values shift with scene content.","fun_headline_variants_meta":{"raw":{"variants":["One-shot depth alignment gives robots metric depth from RGB","MOMA: single RGB image to metric depth in one shot","Scale-rotation-shift alignment unlocks metric depth from one RGB","No depth sensor: one-shot RGB metric depth for grasping","MOMA converts a single RGB image into grasp-ready metric depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1550,"prompt_tokens":992,"completion_tokens":558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":475}},"tokens_in":608,"tokens_out":558,"duration_ms":5751,"temperature":1.0,"reasoning_tokens":475,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:12:11.779801+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the aligned depth error on object masks across many scene variations using a single calibration, and compare with re-fitting the transform per frame; a large gap would mean the one-shot transform does not transfer to novel objects.","supporting_citations":[{"cited_title":"Towards robust monocular depth estimation: Mixing datasets for zero- shot cross-dataset transfer,","cited_arxiv_id":null,"evidence_quote":"Supplies the median-based normalization variant and is cited for global scale-shift fitting."},{"cited_title":"Toward 3d scene reconstruction from locally scale-aligned monocular video depth,","cited_arxiv_id":null,"evidence_quote":"Supplies the locally weighted linear regression alignment baseline."},{"cited_title":"Garc ´ıa, J","cited_arxiv_id":null,"evidence_quote":"Supplies the min-max normalization that precedes the alignment fit."}],"review_version":2}