{"id":"0f8a2fca-31b8-4d6a-94fa-a19aca5ec36a","arxiv_id":"2509.05599","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Monocular glass detection improves when depth is formulated as 3D plane regression and trained with a plane distance loss on a new real-world dataset of complex glass scenes.","lead":"These researchers built a real-world dataset of glass surfaces with 3D annotations and trained a network that detects glass regions and estimates their depth from a single camera image. The method outperforms existing glass segmentation and depth estimation systems, which matters for robots navigating and manipulating around transparent surfaces.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth labels are generated by fitting planes to manually placed 3D boxes on a LiDAR map (Sec. III-C), and Sec.","rationale":"Reader identified the same weakest assumption, and the manuscript itself flags it in Sec. V-F. The central novelty of MonoGlass3D is plane regression for 3D glass localization; the empirical claim that it beats direct depth regression rests entirely on the quality of the plane-projected depth labels. If those labels are not grounded in actual range measurements on glass, then the evaluation is self-referential: the proposed model is designed to produce exactly the representation used to synthesize the ground truth, while the direct-regression baselines are compared against a planar target that may not match physical surfaces. The held-out scenes in Table V address training/validation leakage but not this label-validity question. The segmentation results on GDD and GSD are external and give some support to the architecture, but they do not validate the 3D claim. A physical measurement test is feasible with modest effort, since the dataset is not yet released. I therefore keep the reader's CONDITIONAL verdict: the concern is specific, testable, and currently unresolved, but not evidence of a negative result.","tokens_in":14701,"tokens_out":10963,"duration_ms":103654,"concrete_test":"Physically verify a sample of the depth labels: for at least 20 glass panes spread over 5 or more dataset scenes (including day/night and indoor/outdoor), measure the metric distance from the camera origin to the glass surface at several pixel positions by placing a small reflective target on the glass or using a laser rangefinder at near-normal incidence. Compare those measurements with the Eq. 4 depth values at the same pixels, computing mean absolute error and the error of the fitted plane normal. If the mean error is comparable to or larger than the MonoGlass3D advantage over Depth Anything V2 in Table V (Abs Rel 0.005, MAE about 0.003 m), the plane-fit labels are biased and the central depth claim is not established; if the errors are at the millimeter-to-low-centimeter level, the labels and the evaluation are validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The depth ground truth used for both training and evaluation is constructed in Sec. III-C, not measured: annotators place thin 3D bounding boxes on a FAST-LIVO2 point-cloud map, the box vertices are fit to a plane via Eq. 3, and every pixel under a 2D mask is projected onto that plane via Eq. 4. The accuracy of these labels thus depends on whether the point cloud contains reliable 3D returns on the glass surface. The LiDAR is deliberately operated in single-first-return mode, and Sec. V-F states that 'the 3D positions of glass are not directly observable.' On transparent or reflective panes, first returns may come from background objects, from the glass frame, or from refracted paths, so the fitted plane is at least partly a manual inference rather than a range measurement. Since MonoGlass3D outputs plane parameters and is trained and evaluated on exactly these plane-projected depths, its margin over direct depth-regression baselines (Table I: 0.063 vs. 0.067 Abs Rel; Table V: 0.083 vs. 0.088 Abs Rel) could reflect structural alignment with the label-generation model instead of physical localization accuracy. The held-out scene split in Table V does not remove this concern because those labels inherit the same plane-fit assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents MonoGlass3D, a monocular 3D glass detection method, together with a new real-world dataset. The dataset contains 1,437 annotated frames from 50 scenes, with glass segmentation masks, 3D plane parameters, and depth maps. Depth labels are generated by manually placing thin 3D bounding boxes on a FAST-LIVO2 point-cloud map, fitting a plane to each box, and projecting 2D masks onto those planes. The proposed network uses a DINOv2 backbone with a centerness-based adaptive feature fusion module, a plane regression head that predicts angles and intercept, and attention layers, trained with segmentation, centerness, and plane losses. Experiments report state-of-the-art glass segmentation on GDD and GSD, and the best depth estimation accuracy on the new dataset and on a held-out scene split, while using fewer parameters than the largest baselines. An ablation study isolates the contributions of adaptive fusion, attention, cascades, and the proposed plane distance loss.","tokens_in":14933,"tokens_out":7067,"duration_ms":64920,"significance":"If the depth ground truth is trustworthy, the dataset is a useful contribution for robotic perception of transparent surfaces, and the plane regression formulation with a viewpoint-invariant plane distance loss is a well-motivated alternative to direct depth regression. The segmentation results are validated on external public benchmarks (GDD and GSD), which lends credibility to the architecture's representational strength. The paper also provides clear ablations and promises code and dataset release. However, the significance of the core depth claims is conditional on the validity of the synthesized plane-based depth labels, and the current manuscript does not provide independent evidence for that validity.","major_comments":[{"comment":"The depth ground truth is not a direct measurement: it is constructed by fitting planes to 3D boxes manually placed on a LiDAR point-cloud map, and Section V-F concedes that 'the 3D positions of glass are not directly observable'. With the LiDAR operated in single-first-return mode, returns may originate from objects behind the glass, from the frame, or from refracted paths, so the fitted plane is an annotation assumption rather than a range measurement. Because MonoGlass3D outputs plane parameters and is trained and evaluated on depths projected from exactly these fitted planes, the depth margins over baselines in Tables I and V (e.g., 0.063 vs. 0.067 Abs Rel on the full set) could reflect structural alignment with the label-generation model rather than physical localization accuracy. I request an independent validation of a sample of depth labels, for example by comparing predicted plane distances against hand-measured distances or known architectural plans, or by providing point-cloud evidence that first returns lie on the glass surface itself.","section":"Section III-C, Eqs. (1)-(4); Section V-F"},{"comment":"The GDD and GSD segmentation comparisons are performed with plane regression layers disabled, yet the text presents these results as 'our model achieves state-of-the-art performance' and attributes them to MonoGlass3D. Table IV shows that the full architecture with cascades and plane losses changes IoU on the in-house split (e.g., IoU 0.925 without cascades vs. 0.942 with them). The reader cannot tell whether the segmentation gains come from the full MonoGlass3D design or from a segmentation-only variant. Please state explicitly which components are disabled for these experiments and justify that the variant is representative of the proposed method; otherwise the external segmentation claim tests a different model.","section":"Section V-C.2, Tables II and III"},{"comment":"The angular parameterization appears not to cover all plane normals in the -z hemisphere. For a normal with z<0 and x<0, Eq. (5) gives θ2 = -arccos(z/r_xz); as z/r_xz approaches -1, this value approaches -π, which is outside the claimed range [-π/2, π/2] that the text uses for tanh up-scaling. The case r_xz=0 is also undefined, which would affect horizontal glass surfaces such as ceilings or floors. This means certain plane orientations cannot be represented by the network's output range. Please clarify the intended angular range, correct Eq. (5) if needed, and discuss singularity handling.","section":"Section IV-B, Eq. (5)"},{"comment":"The held-out scene experiment uses six scenes from the same dataset, and their labels inherit the same plane-fit assumption as the training data. As a generalization test, this setup cannot detect systematic bias in the label-generation process; it only checks generalization to new scenes under the same annotation protocol. Please add results on an independently measured scene (e.g., with a different depth sensor or manual distance measurements) or, at minimum, provide a quantitative analysis of label uncertainty across the annotation pipeline.","section":"Section V-E, Table V"}],"minor_comments":[{"comment":"The abstract claims 'precise 3D annotations', but Section V-F states that 3D positions of glass are not directly observable; please temper the abstract or clarify that precision is relative to the plane-fit assumption.","section":"Abstract and Section V-F"},{"comment":"There is a typo in the sentence 'GWDepth [11] uses a depth camera to capture the depth values at glass boundaries and and interpolate pixel-wise glass depths' – 'and and' should be 'and'.","section":"Section II-B"},{"comment":"The term 'interception' should be 'intercept' throughout the description of plane parameters d.","section":"Section III-C, Eq. (3)"},{"comment":"The notation uses r both as the scalar magnitude (r=1 for unit vectors) and as part of r_xz; please use distinct symbols to avoid confusion.","section":"Section IV-B, Eq. (5)"},{"comment":"The loss term is denoted 'lp' in Eq. (8) but 'L_p' elsewhere; unify the notation.","section":"Section IV-D, Eq. (8)"},{"comment":"The ablation table is difficult to read because the checkmark columns are not labelled clearly for each row; please reformat it so that each configuration is unambiguous.","section":"Table IV"},{"comment":"The phrase '52.4 Mtrainable parameters' is missing a space; it should read '52.4 M trainable parameters'.","section":"Section V-C.1"}],"recommendation":"major_revision","confidential_remarks":"The core concern is the validity of the depth ground truth. The external segmentation results are credible, but the depth claims rest on labels that are inferred rather than measured, and the manuscript itself acknowledges that glass positions are not directly observable. If the authors can provide independent validation of even a subset of labels, the paper could be publishable; without it, the depth comparison should be framed as a comparison under the authors' annotation model. I also recommend checking the angular parameterization in Eq. (5) carefully before accepting the plane regression claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read MonoGlass3D. The headline: it's a solid engineering paper with a genuinely useful new dataset, but the depth evaluation is built on labels that are plane fits to manual boxes, not measured glass depths, so the main depth claims should be read as that — comparisons on a plane-fit benchmark, not physical localization accuracy.\n\nWhat's new and good: the dataset is the first to my knowledge with multiple glass planes per image (up to 10) and night scenes, and the annotation pipeline is practical. The method is also thoughtful: polar parameterization for plane normals reduces the output space, the plane distance loss is an original way to supervise geometry without the angle-vs-depth error imbalance, and the centerness-based adaptive fusion is simple and shows clear ablative gains. The external segmentation results on GDD and GSD are credible — they test on separate benchmarks and beat bigger models with a 34M-parameter variant. The held-out scene experiment (six scenes absent from training) is a good idea, and segmentation there stays strong.\n\nWhere I'd push back: the depth ground truth in Sec. III-C is constructed, not measured. Planes are fit to thin bounding boxes manually placed on a FAST-LIVO2 point cloud map, and the LiDAR runs single-first-return mode. On transparent glass, first returns can come from background or the frame; the paper itself says in Sec. V-F that the 3D positions of glass are not directly observable. That means the labels encode an annotator's geometric inference, and a method trained to regress planes is structurally aligned with the label model. Table I's margins (0.063 vs 0.067 Abs Rel) are small enough that this alignment could explain part of the gap. Also, the main validation appears to be an image-level split — the paper doesn't say scenes were separated for the main Table I, only for the six-scene experiment. If training and validation share scenes, the plane parameters for those glass panes may have been seen before. That's a bigger concern than the reader's conditional verdict suggests.\n\nThe code and dataset are promised but not released, which blocks verification. I'd want to see the artifacts, plus an analysis of how often the point cloud actually has returns on glass vs. background, before trusting the depth numbers.\n\nIn sum: the segmentation work and the dataset are valuable; the depth evaluation is a benchmark artifact that needs explicit acknowledgment. I'd send it to review, but the reviewers should demand the scene-split clarification, the artifact release, and a sensor-level validation of the label process.","headline":"Solid dataset and method, but the depth labels are plane fits to manually placed boxes, not measured glass depths, so the depth claims are weaker than the segmentation ones.","tokens_in":15455,"tokens_out":2843,"would_cite":true,"duration_ms":25145,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Glass is nearly invisible to depth sensors because light passes through it; MonoGlass3D claims the fix is to stop estimating depth pixel-by-pixel and instead regress the 3D plane each glass surface lies on, backed by a new 1,437-frame…","keywords":["glass detection","monocular 3D detection","plane parameter regression","depth estimation","adaptive feature fusion","centerness","plane distance loss","transparent surfaces"],"falsifier":"Take a handful of new scenes with glass, measure the true glass-surface positions independently of LiDAR-through-glass — for example with a structured-light scan, a rangefinder placed flat against the pane, or direct tape measurement — and compare MonoGlass3D's predicted planes and its training labels against those measurements. If the plane-fitted labels and the network's predictions agree with each other but disagree with the independent measurements by the same margin as the reported errors, then the plane-fit labeling pipeline, not the network, is responsible for the good numbers.","tokens_in":14523,"feed_emoji":"🪟","tokens_out":11365,"duration_ms":91826,"temperature":0.7,"pith_summary":"The paper is trying to establish that monocular 3D glass detection should be formulated as plane parameter regression with centerness-guided adaptive fusion, rather than as direct pixel-by-pixel depth estimation. It contributes a new real-world dataset of 1,437 annotated frames from 50 scenes, with up to ten glass planes per image, depth ranges from 0.07 m to 17.76 m, and 21 night scenes, where depth labels come from projecting 2D masks onto planes fitted to LiDAR-derived 3D boxes. On this dataset, MonoGlass3D reports the best depth metrics among the compared methods, including general-purpose monocular depth models, using 52.4M parameters, and it reports state-of-the-art glass segmentation on the GDD and GSD benchmarks when the plane branch is switched off. A sympathetic reader would take the core claim to be that the planarity of glass is a stronger cue than any appearance-based context, and that exploiting it explicitly is what makes transparent surfaces tractable for a single RGB camera.","feed_headline":"Glass depth from plane regression beats direct estimation","feed_subtitle":"A new real-world dataset and a 52.4M-parameter network bring transparent surfaces into monocular 3D perception.","key_machinery":"The load-bearing mechanism is the plane parameterization of glass. Each glass surface is represented by $(\\theta_1, \\theta_2, d)$, where $\\theta_1,\\theta_2$ are polar angles of the unit normal restricted to the $-z$ hemisphere and $d$ is the plane intercept; this reduces the natural four-variable plane representation to three parameters and removes the need for post-hoc normalization. Predicted plane parameters are converted into depth by projecting each pixel onto its plane, $depth = d / (n^T K^{-1}[u_x, u_y, 1]^T)$, so depth inherits the geometric regularity of the plane. Training is driven by a plane distance loss, which places four equally spaced points around each predicted pixel's 3D position and sums their distances to the ground-truth plane; unlike a depth loss, this error is invariant under viewpoint and plane orientation. The second mechanism is centerness-based adaptive fusion: centerness $C=\\sqrt{d_{\\min}/d_{\\max}}$ is predicted early, and features entering the segmentation and plane heads are multiplied by $(1-C)$, steering the network toward the context-rich region around glass boundaries; the cascade repeats this at three backbone scales, with self- and cross-attention refining the final plane and segmentation features.","core_discovery":"On the paper's own terms, the central discovery is that the transparency of glass stops being the main obstacle once the network is constrained to see glass as 3D planes. The paper shows that regressing three plane parameters — two polar angles for the normal and one intercept — and projecting them to depth yields flatter, more accurate depth on glass surfaces than direct depth regression, and that reweighting features by $1-C$, where $C=\\sqrt{d_{\\min}/d_{\\max}}$ is a per-pixel centerness, provides the contextual signal that appearance cues like boundaries and reflections cannot. The supporting evidence is a new dataset whose annotation pipeline produces dense depth labels from sparse 3D plane fits, and evaluations in which MonoGlass3D achieves the best reported depth error on all metrics, with IoU of 0.920 on GDD and 0.872 on GSD for segmentation and stronger held-out-scene depth than the baselines.","pith_inferences":["Our inference: the same plane-regression recipe should transfer to other planar transparent or reflective surfaces — mirrors, display cases, acrylic partitions — because they share the same failure mode of letting light and depth-sensor signals pass through; the paper does not test these.","Our inference: the plane distance loss could work as a geometric regularizer for general monocular depth estimation in piecewise-planar scenes, not just glass; the paper's ablation shows it beats depth supervision in this architecture, but the broader transfer is unverified.","Our inference: fusing the predicted plane intercepts with sparse LiDAR returns that do hit glass (available in some frames) could resolve monocular scale ambiguity and is a direct, testable extension the paper leaves for future work.","Our inference: the $(1-C)$ centerness weighting may benefit detection of any object whose interior is less informative than its boundary — windows, mirrors, fences — which would make the mechanism useful beyond glass."],"forward_implications":["A robot with only an RGB camera could obtain glass depth accurate enough for navigation and manipulation, sidestepping the LiDAR and depth cameras that see through or ignore transparent surfaces.","Segmentation and 3D localization collapse into one forward pass: the same features produce masks and plane parameters, so 'where the glass is' and 'where it sits in 3D' arrive together.","Coplanar glass instances — several windows on one wall — will be reconstructed as mutually consistent flat surfaces because they share plane parameters, removing the noisy interior depths that direct regression produces on featureless glass.","Since the plane distance loss is orientation-invariant, training is not dominated by frontoparallel planes, and oblique or multi-angle glass configurations contribute equally to the gradient.","The new dataset, with up to ten planes per image and night-time illumination, raises the ceiling for evaluating 3D glass perception beyond the one-or-two-plane datasets used previously."],"supporting_citations":[{"why":"Provides the previous 3D glass dataset and method whose limited plane diversity motivates the new dataset.","marker":"[11]"},{"why":"Serves as the main depth-estimation baseline and supplies the backbone initialization for MonoGlass3D.","marker":"[12]"},{"why":"Serves as the second depth-estimation baseline, fine-tuned on the new dataset for comparison.","marker":"[31]"},{"why":"Generates the colored LiDAR-visual point-cloud maps from which the 3D glass annotations are built.","marker":"[28]"},{"why":"Defines the centerness measure that the adaptive fusion module reuses for feature reweighting.","marker":"[15]"},{"why":"Establishes the plane-parameter regression paradigm that the paper adapts to glass surfaces.","marker":"[24]"},{"why":"Supplies the piecewise-planarity prior that justifies replacing depth regression with plane projection.","marker":"[27]"},{"why":"Defines the GDD glass-segmentation benchmark and the GDNet baseline compared in Table II.","marker":"[2]"},{"why":"Defines the GSD glass-segmentation benchmark and the baseline compared in Table III.","marker":"[17]"}],"fun_headline_variants":["Plane regression sharpens monocular glass depth","New dataset and plane fitting for glass 3D","Monocular glass depth: geometry beats appearance","Adaptive fusion and plane priors for glass detection","Regressing planes for transparent surface 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The depth labels that train and evaluate everything are produced by fitting planes to manual 3D boxes on a LiDAR-visual SLAM point cloud, which assumes the point cloud actually contains points on the glass surface rather than behind it — an assumption the paper itself concedes is not directly observable.","fun_headline_variants_meta":{"raw":{"variants":["Plane regression sharpens monocular glass depth","New dataset and plane fitting for glass 3D","Monocular glass depth: geometry beats appearance","Adaptive fusion and plane priors for glass detection","Regressing planes for transparent surface 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000603,"raw_usage":{"total_tokens":2806,"prompt_tokens":930,"completion_tokens":1876,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":1804}},"tokens_in":546,"tokens_out":1876,"duration_ms":14527,"temperature":1.0,"reasoning_tokens":1804,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:21:53.099816+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a handful of new scenes with glass, measure the true glass-surface positions independently of LiDAR-through-glass — for example with a structured-light scan, a rangefinder placed flat against the pane, or direct tape measurement — and compare MonoGlass3D's predicted planes and its training labels against those measurements. If the plane-fitted labels and the network's predictions agree with each other but disagree with the independent measurements by the same margin as the reported errors, then the plane-fit labeling pipeline, not the network, is responsible for the good numbers.","supporting_citations":[{"cited_title":"Monocular depth estimation for glass walls with context: a new dataset and method,","cited_arxiv_id":null,"evidence_quote":"Provides the previous 3D glass dataset and method whose limited plane diversity motivates the new dataset."},{"cited_title":"FAST-LIVO2: Fast, direct lidar-inertial-visual odometry,","cited_arxiv_id":null,"evidence_quote":"Generates the colored LiDAR-visual point-cloud maps from which the 3D glass annotations are built."},{"cited_title":"Polarmask: Single shot instance segmentation with polar rep- resentation,","cited_arxiv_id":null,"evidence_quote":"Defines the centerness measure that the adaptive fusion module reuses for feature reweighting."},{"cited_title":"PlaneNet: Piece-wise planar reconstruction from a single RGB image,","cited_arxiv_id":null,"evidence_quote":"Establishes the plane-parameter regression paradigm that the paper adapts to glass surfaces."},{"cited_title":"P3Depth: Monocular depth estimation with a piecewise planarity prior,","cited_arxiv_id":null,"evidence_quote":"Supplies the piecewise-planarity prior that justifies replacing depth regression with plane projection."},{"cited_title":"Don’t hit me! glass detection in real-world scenes,","cited_arxiv_id":null,"evidence_quote":"Defines the GDD glass-segmentation benchmark and the GDNet baseline compared in Table II."},{"cited_title":"Rich context aggregation with reflection prior for glass surface detection,","cited_arxiv_id":null,"evidence_quote":"Defines the GSD glass-segmentation benchmark and the baseline compared in Table III."}],"review_version":2}