{"id":"fbb6d78a-4f6c-49e8-88e6-dbbbad6604ad","arxiv_id":"2509.08159","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A monotonic spline fitted to sparse visual-inertial features rescales monocular relative depth to metric depth, enabling collision avoidance with a single camera and IMU.","lead":"A drone estimates real-world distances to obstacles using just one camera and a motion sensor, by correcting a neural network's rough depth guesses with a handful of measured 3D points. This makes lightweight, low-cost quadrotors capable of avoiding collisions in unknown places without LiDAR or stereo cameras.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation metric in Eq. (3) is non-standard and GT baselines fail in Table I, undermining the quantitative evidence for the central claim.","rationale":"The reader's weakest_assumption centers on the generalization of the sparse-feature rescaling, which is a valid concern and is explicitly admitted in the Limitations section. However, the most load-bearing issue for the central claim is the evaluation metric itself: Eq. (3) is non-standard and produces absurd GT-baseline results in Table I. This directly undermines the quantitative evidence that the proposed method yields accurate metric depth, which is the foundation of the contribution. If the AbsRel metric is flawed or the evaluation code has a bug, the paper's reported accuracy numbers and the selection of the monotonic spline as the best method are unreliable. The real-world navigation success provides some support for the system as a whole, but it is only two trials and does not isolate depth accuracy. Therefore, the paper should remain CONDITIONAL: the central idea may be sound, but the evaluation must be corrected before the quantitative claims can be accepted. I partially agree with the reader because they noted the GT failures as an issue, but they did not identify it as the weakest assumption; I argue it is more load-bearing than the sparse-feature extrapolation, which is an acknowledged limitation rather than a hidden flaw. My recommendation is UNCHANGED because the reader's verdict of CONDITIONAL is already appropriate given this concern.","tokens_in":10333,"tokens_out":9498,"duration_ms":105424,"concrete_test":"Recompute Table I using the standard AbsRel definition, |zpred - zgt| / zgt, and verify that the GT baseline outperforms the VINS-based fit for every rescaling method in every environment. Specifically, reproduce the Drone Dome degree-1 polynomial entry: if GT AbsRel remains higher than VINS AbsRel after this correction, inspect the fitting/extrapolation code for a bug (e.g., disparity vs. depth conversion, clipping, or column misalignment). Also report δ1 and per-environment error bars.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that monotonic spline rescaling of DepthAnythingV2 relative depth yields accurate metric depth for navigation—rests on the quantitative evaluation in Table I. However, Eq. (3) defines AbsRel as (1/N) * sum |zpred - zgt| / zpred, normalizing by the predicted depth rather than the ground truth. This is non-standard (the usual metric normalizes by ground truth) and systematically penalizes underestimates of depth (small zpred) far more than overestimates, which is the opposite of what obstacle avoidance needs: predicting an obstacle as farther than it is (large zpred) is dangerous yet yields small AbsRel. This metric choice appears to produce the counterintuitive GT-baseline failures: in Table I, the 'GT' columns, which rescale using ground-truth depths and are expected to outperform the proposed VINS-based rescaling, often show worse AbsRel (e.g., Drone Dome deg-1 poly GT AbsRel = 2.613 vs. VINS = 0.894; weighted average deg-1 GT = 1.081 vs. VINS = 0.436). Such results are implausible if the evaluation harness is correct, and suggest either a bug in the fitting/evaluation code or a mislabeling of columns. Since these numbers are the primary evidence for the method's metric accuracy, the evaluation harness is a load-bearing weakness. The paper's own limitation about sparse-feature extrapolation (Section V) is an acknowledged scope restriction, not an internal contradiction, so it is less damaging than a flawed evaluation metric.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a zero-shot method for metric depth estimation from a monocular camera and IMU. A VINS sliding-window optimizer produces a sparse 3D feature map; the relative disparity output of DepthAnythingV2 is rescaled into metric disparity by fitting one of several functions (polynomial, exponential, smoothing spline, monotonic spline) to the sparse feature correspondences. The monotonic spline version is selected after simulation ablations, integrated with a motion-primitive planner, and deployed on a quadrotor in a dusty tunnel, where it reportedly runs at 15 Hz and enables collision avoidance. The manuscript also releases software and includes a real-world hardware validation.","tokens_in":10731,"tokens_out":5219,"duration_ms":65405,"significance":"If the reported results are reliable, this is a practical contribution: it offers a lightweight, fine-tuning-free way to obtain metric depth for MAV navigation using sensors already present on many platforms, with a real hardware demonstration in a challenging environment. The open-source release and the careful treatment of real-world operational constraints (dust, compute limits, feature scarcity) are strengths. However, the quantitative evaluation currently contains a nonstandard error metric and implausible ground-truth baseline results, so the central accuracy claim is not yet supported as written.","major_comments":[{"comment":"The paper defines AbsRel as (1/N)Σ|z_pred−z_gt|/z_pred, while the surrounding text says it is normalized by ground-truth depth. The standard absolute relative error normalizes by z_gt. With the printed denominator, overestimates (z_pred>z_gt) are discounted and underestimates are heavily penalized; for obstacle avoidance this is the wrong asymmetry, because an overestimate is the dangerous case. Since Table I is the primary quantitative evidence, the evaluation should be redone with the standard metric or the choice must be strongly justified.","section":"§IV.B, Eq. (3)"},{"comment":"The columns labeled GT are described as an expected upper bound because they fit the rescaling using ground-truth sparse depths. Yet GT often performs much worse than the VINS-based rescaling: Drone Dome deg-1 polynomial GT AbsRel=2.613 vs. VINS=0.894, and deg-3 GT=20.259 vs. VINS=0.457; the weighted-average deg-1 GT=1.081 vs. VINS=0.436. This is implausible under a correct evaluation pipeline. It may result from the nonstandard metric in Eq. (3), or from a fitting/evaluation bug. Either way, the Table I results must be corrected and re-reported before the quantitative comparison can be trusted.","section":"Table I"},{"comment":"The rescaling pipeline depends on several hand-set hyperparameters (spline knot count, smoothing S, penalties λ and κ, minimum sparse-feature count) with no sensitivity study, so the reported monotonic-spline performance is tied to those choices. More importantly, the Limitation section admits that depth estimates are unreliable outside the depth range spanned by the sampled sparse features. Since this is a core premise of the method, the manuscript should quantify how often and how severely this occurs in the test environments and should temper the 'zero-shot' claim accordingly.","section":"§III.C and §V"}],"minor_comments":[{"comment":"The δ1 condition is written as two separate inequalities; this is equivalent to the standard max-ratio condition, but the notation should be clarified. Table headers use 'Delta1' rather than 'δ1'.","section":"§IV.B, Eq. (4)"},{"comment":"The notation 'C 2 spline' should read 'C^2 spline'; 'preceeding' should be 'preceding'. Please polish these and similar typos.","section":"§III.B.3 and §III.B.5"},{"comment":"Please clarify what the percentages represent (e.g., fraction of trials) and what 'Total' aggregates across the three sewer maps. The current bar chart is hard to interpret without this detail.","section":"Fig. 4"},{"comment":"State explicitly whether the metrics in Table I are computed over all valid image pixels or only at the sparse feature pixels used for fitting. This distinction is important for interpreting the rescaling accuracy and for comparing with prior work such as [10].","section":"§IV.C"}],"recommendation":"major_revision","confidential_remarks":"The hardware demonstration is valuable and the core formulation is sound, but the quantitative backbone of the paper—Table I and Eq. (3)—is currently unreliable. I believe a revision that re-runs the evaluation with the standard AbsRel metric and reconciles the ground-truth baseline could make the paper acceptable. The authors should also be asked to clarify the evaluation domain (full image vs. sparse feature pixels) and provide a sensitivity analysis for the spline hyperparameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's the take on arXiv:2509.08159. The core idea: take DepthAnythingV2 relative depth, fit a monotonic spline mapping relative disparity to metric disparity using sparse 3D features from a sliding-window VINS, and use the rescaled depth for collision avoidance. That combination—monotonic spline with an explicit non-monotonicity penalty, evaluated against several simpler function families, and deployed on a real quadrotor at 15 Hz—is new relative to Marsal et al., who used linear regression on ground-truth features. The hardware demonstration in a dusty tunnel with a compute-limited Orin AGX is the strongest part; it shows the pipeline can actually close the loop.\n\nThe paper is honest about limitations: it says depth outside the feature range is unreliable, and it notes chattering. That is a real scope restriction, not hidden.\n\nSoft spots. The biggest is the evaluation metric, Eq. (3). AbsRel is defined with predicted depth in the denominator, not ground truth. Non-standard. It makes overestimates of depth look cheap, which is the wrong direction for obstacle avoidance—if you think an obstacle is farther than it is, you get a small error and then fly into it. The standard metric divides by ground truth. This likely explains the implausible Table I results where the GT baseline (rescaling with ground-truth depth) looks worse than noisy VINS-based rescaling. I cannot accept those numbers until the authors recompute with standard AbsRel and explain why GT underperforms. Could be a bug or column mislabel, but as printed it undermines confidence.\n\nAlso, navigation results in Fig. 4 have no error bars and 5 trials per pair; fine for a demo, but not statistical evidence. And there are hand-set hyperparameters (spline knots, penalties) with no sensitivity analysis. Minor.\n\nThe central claim—that you can get usable metric depth for navigation by rescaling relative MDE with VINS features—is plausible and supported by the hardware run, even if simulation evidence is shaky. The paper deserves a serious referee, but I would want them to fix the metric, rerun Table I, and report standard AbsRel with error bars. If GT baseline still underperforms, the harness needs a harder look.\n\nFor you: worth bringing to a reading group if you work on MAV perception or test-time adaptation. I probably would not cite the numbers until rerun, but the system integration is a useful data point.\n\nRecommendation: send to peer review, with the evaluation metric as the main ask. Serious thinker: yes.","headline":"Solid MAV depth-rescaling paper with a real hardware demo; the AbsRel metric in Eq. (3) is non-standard and the GT baseline failures in Table I need explanation before trusting the quantitative claims.","tokens_in":11199,"tokens_out":2989,"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":"The paper claims that a drone with only one camera and an IMU can estimate metric depth for obstacle avoidance by fitting a monotonic spline that maps relative depth predictions to sparse 3D feature points from a visual-inertial navigation","keywords":["metric depth estimation","monocular visual-inertial navigation","zero-shot rescaling","relative depth","collision avoidance","quadrotor","monotonic spline","sparse 3D features"],"falsifier":"Fly in a straight line toward a flat wall in a low-texture corridor while the tracked VINS features are all on nearby floor texture; record ground-truth distance to the wall with a rangefinder or motion capture and compare the rescaled depth at the wall's pixels. If the predicted wall distance deviates from the true distance by more than the planner's tolerance at the moment the wall enters the image, the assumption that the spline extrapolates beyond the feature range fails.","tokens_in":10250,"feed_emoji":"🚁","tokens_out":6088,"duration_ms":68696,"temperature":0.7,"pith_summary":"The paper sets out to show that a quadrotor can navigate cluttered, a priori unknown environments using only a monocular camera and an IMU, without heavy depth sensors or environment-specific fine-tuning. It does this by taking relative depth images from a pretrained monocular depth network and rescaling them into metric depth using a monotonic spline fitted to the sparse 3D feature map that a visual-inertial navigation system already produces. The central claim is that this online rescaling generalizes across the image well enough to support collision avoidance, with depth estimates running onboard at 15 Hz on a compute-constrained quadrotor. If true, this provides a zero-shot, sensor-light path to metric depth for small aerial robots in unknown environments.","feed_headline":"Spline trick turns monocular depth into metric depth at 15 Hz","feed_subtitle":"Zero-shot method fits a monotonic spline to sparse VINS 3D features, letting a small quadrotor avoid obstacles without fine-tuning.","key_machinery":"The central object is the monotonic spline: a cubic B-spline function fitted by penalized least squares to pairs of relative disparity (from the depth network) and metric disparity (from the VINS sparse feature map). Monotonicity is enforced through a penalty on negative first differences, so the fitted curve preserves the ordering of depths. The spline is fitted online per frame from the sparse feature set and then applied densely to the whole image, which is what converts a scale-ambiguous relative depth image into a metric depth image.","core_discovery":"The paper's central claim is that relative monocular depth estimates can be turned into metric depth without retraining by fitting a monotonic spline from the sparse 3D feature map produced by a sliding-window visual-inertial navigation system to the relative disparity values in the same image. The fitted function is applied to every pixel of the depth image, and the result—metric depth in meters—is used directly by a motion-primitives planner for collision avoidance. In simulation, the monotonic spline is the most consistent rescaling strategy across both confined and open environments, and on hardware it enabled a quadrotor with only a forward camera and IMU to complete two obstacle-avoida","pith_inferences":["The same spline-fitting procedure should transfer to other relative-depth networks, because it only sees disparity pairs and does not depend on the internals of the specific depth network used.","A safety-focused planner could treat pixels whose relative disparity lies outside the range covered by current VINS features as unknown, since the paper's limitation section states that extrapolation beyond the sampled feature range is unreliable.","Weighting the spline fit by VINS feature covariance would be a natural next step; the paper mentions uncertainty-aware rescaling as future work, and simulation benchmarks could quantify the benefit in low-texture scenes."],"forward_implications":["Aerial robots can run collision avoidance with only a monocular camera and IMU, since the metric depth image is produced onboard at 15 Hz and feeds the planner directly.","No environment-specific fine-tuning or ground-truth depth is needed at deployment; the rescaling is computed on the fly from the VINS feature map.","The method shows consistent accuracy in both confined and open simulation environments, with the monotonic spline performing best across both.","Existing planners and costmaps that assume metric depth can be used unchanged, since the output is a dense metric image rather than relative depth."],"supporting_citations":[{"why":"Depth Anything V2 supplies the relative depth image that must be rescaled; the pipeline begins from these predictions.","marker":"[6]"},{"why":"The sliding-window visual-inertial optimizer that produces the sparse 3D feature map used as metric ground truth for fitting the spline.","marker":"[18]"},{"why":"The most similar prior method, which rescales relative depth by linear regression to sparse points; the paper contrasts its use of ground-truth-derived points with real VINS features.","marker":"[10]"},{"why":"The motion-primitives planner used for navigation experiments, demonstrating that the rescaled metric depth supports collision avoidance.","marker":"[27]"},{"why":"Flightmare simulator used to collect the photo-realistic mine, sewer, and drone-dome evaluation datasets.","marker":"[25]"},{"why":"Unimodal smoothing formulation that supplies the monotonicity penalty used in the monotonic spline rescaling.","marker":"[23]"},{"why":"B-spline basis definition used to construct the spline function.","marker":"[24]"},{"why":"Spline-fitting algorithm used for the smoothing-spline baseline variants.","marker":"[22]"},{"why":"Prior work that rescales Depth Anything V2 with an active stereo depth sensor; the paper's approach replaces the depth sensor with VINS features.","marker":"[7]"}],"fun_headline_variants":["Spline fit from IMU features makes monocular depth metric at 15 Hz","Zero-shot metric depth via spline on sparse VINS features","Monocular + IMU spline rescaling enables drone obstacle avoidance","15 Hz metric depth from monocular camera with spline rescaling","No retraining: spline fits IMU features to make monocular depth metric"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The rescaling curve is fitted only where the visual-inertial system has metrically accurate 3D features, and those features are assumed accurate and spread widely enough to represent the whole scene, including depths beyond their range.","fun_headline_variants_meta":{"raw":{"variants":["Spline fit from IMU features makes monocular depth metric at 15 Hz","Zero-shot metric depth via spline on sparse VINS features","Monocular + IMU spline rescaling enables drone obstacle avoidance","15 Hz metric depth from monocular camera with spline rescaling","No retraining: spline fits IMU features to make monocular depth metric"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000392,"raw_usage":{"total_tokens":1869,"prompt_tokens":687,"completion_tokens":1182,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":1100}},"tokens_in":431,"tokens_out":1182,"duration_ms":11495,"temperature":1.0,"reasoning_tokens":1100,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:08:41.751555+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fly in a straight line toward a flat wall in a low-texture corridor while the tracked VINS features are all on nearby floor texture; record ground-truth distance to the wall with a rangefinder or motion capture and compare the rescaled depth at the wall's pixels. If the predicted wall distance deviates from the true distance by more than the planner's tolerance at the moment the wall enters the image, the assumption that the spline extrapolates beyond the feature range fails.","supporting_citations":[{"cited_title":"Depth Anything V2,","cited_arxiv_id":null,"evidence_quote":"Depth Anything V2 supplies the relative depth image that must be rescaled; the pipeline begins from these predictions."},{"cited_title":"Resource-constrained state estimation with multi-modal sensing,","cited_arxiv_id":null,"evidence_quote":"The sliding-window visual-inertial optimizer that produces the sparse 3D feature map used as metric ground truth for fitting the spline."},{"cited_title":"Rapid quadrotor navigation in diverse environments using an onboard depth camera,","cited_arxiv_id":null,"evidence_quote":"The motion-primitives planner used for navigation experiments, demonstrating that the rescaled metric depth supports collision avoidance."},{"cited_title":"Flightmare: A flexible quadrotor simulator,","cited_arxiv_id":null,"evidence_quote":"Flightmare simulator used to collect the photo-realistic mine, sewer, and drone-dome evaluation datasets."},{"cited_title":"Unimodal smoothing,","cited_arxiv_id":null,"evidence_quote":"Unimodal smoothing formulation that supplies the monotonicity penalty used in the monotonic spline rescaling."},{"cited_title":"De Boor, A practical guide to splines","cited_arxiv_id":null,"evidence_quote":"B-spline basis definition used to construct the spline function."},{"cited_title":"An algorithm for smoothing, differentiation and in- tegration of experimental data using spline functions,","cited_arxiv_id":null,"evidence_quote":"Spline-fitting algorithm used for the smoothing-spline baseline variants."},{"cited_title":"Reactive Collision Avoidance for Safe Agile Navigation","cited_arxiv_id":"2409.11962","evidence_quote":"Prior work that rescales Depth Anything V2 with an active stereo depth sensor; the paper's approach replaces the depth sensor with VINS features."}],"review_version":1}