{"id":"a938ecb7-0349-48f8-8274-7c66bcdda800","arxiv_id":"2504.12799","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TSGS decouples geometry from appearance in Gaussian splatting and extracts first-surface depth via a transmittance sliding window, improving transparent surface reconstruction on a new synthetic lab-glass benchmark.","lead":"TSGS is a two-stage Gaussian splatting method for transparent surfaces: it first learns geometry from de-lighted images and normal priors, then refines appearance while keeping opacity fixed. It adds a sliding-window first-surface depth extraction and a new TransLab benchmark, reporting large chamfer-distance gains over PGSR on that benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3 mm depth-error bound in Appendix A.3 does not follow from the maximum-weight window heuristic; for high-transparency foregrounds the selected window can sit on the transmitted background rather than the true first surface.","rationale":"The reader's weakest_assumption is precisely the window-containment assumption: that the maximum-weight window actually contains the true first surface. My stress-test confirms and sharpens this into a concrete logical flaw: the Appendix A.3 error bound is invalid unless that containment is guaranteed, and for high-transparency surfaces it is not even the likely case. This is load-bearing because the paper's motivation is millimeter-precise robotic manipulation, and the headline geometric improvement is attributed to the first-surface extraction. However, I do not think this single concern should flip the verdict to REJECT: the empirical TransLab numbers, the DTU results, the ClearPose appendix, and the ablations all provide genuine evidence that the overall TSGS pipeline improves geometry. The missing strong baselines (TranSplat, Car-GS, alpha-NeuS, alpha-surf) and the unreported thresholds remain secondary concerns; they are important for a final ACCEPT but do not themselves invalidate the submitted comparison. The right posture is to keep the reader's CONDITIONAL verdict: the method is promising and likely reproducible, but the core mechanism's guarantee and the claimed error bound need either a corrected derivation or a direct empirical test. My recommended concrete check would settle whether the window heuristic fails exactly where it is needed, and would also force the authors to report the threshold values and the per-pixel depth-error distribution, which are currently missing.","tokens_in":22503,"tokens_out":5988,"duration_ms":68521,"concrete_test":"Use the ground-truth depth maps provided in the TransLab dataset (Appendix A.1). For every pixel with M_T = 1 in a trained TSGS model, record whether the GT first-surface depth lies within the depth span of the selected window W*, and compute |D_first - D_GT|. Stratify these results by foreground transmittance (e.g., by 1 - sum of T_i alpha_i before the first surface) and report the fraction of transparent pixels where GT depth is outside W* or the error exceeds 3 mm. If a meaningful fraction of high-transmittance pixels violates the bound, the Appendix A.3 guarantee fails and the first-surface mechanism is not resolving the dilemma for the cases it targets. A complementary minimal experiment: build a synthetic two-surface scene with a foreground alpha of 0.05 and a bright background, train TSGS, and check whether the selected window W* contains the foreground or the background surface.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central geometric mechanism is the first-surface extraction of Section 3.4. Equation 18 averages Gaussian plane depths inside the window W* that maximizes sum_{i in W*} T_i alpha_i, with W* constrained only by the accumulated-transmittance interval [T_end, T_start]. For a genuinely transparent foreground, the first-surface Gaussians have small alpha_i, so T_i alpha_i at the first surface can be much smaller than T_i alpha_i at a later, more opaque background surface. Nothing in the construction ties the maximum-weight window to the true first surface: the search is over a transmittance range, not over a depth neighborhood, and the assumption that the first significant transmittance drop coincides with the maximum-weight window is empirical rather than derived. Consequently, the Appendix A.3 statement that the depth error is naturally bounded by the window size, at most delta_t = 3 mm, does not follow. The bound is valid only if the true surface depth is inside the selected window W*, which is exactly the unproven assumption. This is not merely a theoretical quibble: the ablation in Table 4 shows that replacing the extraction with standard unbiased depth changes CD only from 1.85 to 1.89 and leaves F1 unchanged, so the empirical evidence that this mechanism is what drives the headline gain is thin. The main benchmark improvement appears to be carried by the normal and de-light priors, not by the window-based extraction itself. I do not dispute the reproducibility of the reported TransLab numbers; the concern is that the paper over-claims the mechanism and the millimeter-level error bound that motivates its robotic-manipulation applications.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"TSGS addresses transparent surface reconstruction with 3D Gaussian Splatting by separating geometry learning from appearance refinement. Stage 1 optimizes geometry with de-lighted hybrid images, normal priors from StableNormal, a transparency-attribute loss, and flatten regularization; Stage 2 fixes opacity and adds anisotropic spherical Gaussian (ASG) appearance modeling supervised by the original images. At inference, a first-surface depth is extracted by restricting the ray to an accumulated-transmittance interval, selecting a sliding window that maximizes the sum of T_i alpha_i, and averaging the Gaussian plane depths in that window. The method is evaluated on a new synthetic TransLab dataset of eight laboratory scenes with ground-truth meshes, on DTU, and on a limited subset of ClearPose; it reports a 37.3% chamfer-distance reduction and an 8.0% F1 improvement over PGSR on TransLab, plus a 0.41 dB PSNR gain.","tokens_in":22776,"tokens_out":4528,"duration_ms":49451,"significance":"The paper targets a real problem, the transparency-depth dilemma in 3DGS, and its proposed pipeline is coherent and reasonably motivated. The authors release code and a dataset, and they include a comparison against baselines augmented with the same normal priors (Table 8), which strengthens the attribution of the gains. The two-stage strategy with frozen opacity is a sensible way to prevent appearance optimization from corrupting geometry, and the use of masked normal and de-light priors is carefully designed. However, the headline geometric improvement is measured on a synthetic dataset, the signature first-surface extraction contributes only a small amount in the ablation (Table 4), and the claimed 3 mm error bound in Appendix A.3 is not actually proven. These issues temper the contribution, although they are addressable within the manuscript's scope.","major_comments":[{"comment":"The statement that the depth error is 'naturally bounded' by the window size, at most delta_t = 3 mm, is not established. The bound is valid only if the true first surface lies inside the selected window W*, but the maximum-weight criterion is a heuristic and is not shown to guarantee this; for a highly transparent foreground with small alpha_i relative to a more opaque background, the maximum-weight window can sit on the transmitted background. Please remove or replace this bound with an empirical error distribution against the ground-truth depth maps available in TransLab, and state explicitly that the assumption of window containment is what the method relies on.","section":"Appendix A.3, Eq. (18)"},{"comment":"The ablation shows that replacing the first-surface extraction with unbiased depth changes chamfer distance only from 1.85 to 1.89 and leaves the F1 score unchanged, a small effect relative to the headline gain over PGSR (1.85 vs. 2.95). This does not support the description of the maximum-weight window extraction as the load-bearing mechanism behind the reported improvement. Please provide a targeted analysis, such as per-transparent-mask errors or a breakdown on the scenes with the strongest transparency effects, and adjust the wording in Sections 1 and 3.4 to reflect the measured contribution.","section":"Table 4, Section 3.4"},{"comment":"Several core hyperparameters are not reported anywhere: T_start, T_end, theta_T, theta_n, and the sliding-window size is given only in the appendix (3 mm). Without these values, without a sensitivity analysis, and without variance across random seeds or initialization, the TransLab results cannot be fully reproduced or assessed for stability. Please report the settings and include error bars or seed variance for the main quantitative tables.","section":"Section 4.1, Section 3.4"},{"comment":"The paper honestly acknowledges the single-layer transparency assumption, but this limitation should be connected to the benchmark claims. TransLab is a synthetic dataset and, as described, does not include the multi-layer refractive cases (e.g., liquid inside a beaker) that motivate the lab-manipulation application, so the 37.3% chamfer-distance improvement may not transfer to those cases. Please scope the claims accordingly and, if feasible, add a TransLab variant with such multi-layer scenes.","section":"Appendix A.4"}],"minor_comments":[{"comment":"There is a typo in the Introduction: 'chamber distance' should be 'chamfer distance'; the same misspelling appears in the abstract-related text and should be checked throughout.","section":"Section 1"},{"comment":"The abstract and Section 4.1 describe TransLab as close to 'realistic conditions,' while Appendix A.1 states that it is a synthetic benchmark rendered with Blender's PBR engine; the main text should state clearly that the dataset is synthetic.","section":"Section 4.1 / Appendix A.1"},{"comment":"In the phrase 'T_i guadually decreases,' 'guadually' should be 'gradually'; please also proofread the equations for formatting issues such as the garbled subscripts in Eq. (5) and the surrounding text.","section":"Section 3.4"},{"comment":"The ClearPose evaluation uses only one scene per set, subsamples one frame every 100, and reports a unidirectional chamfer distance; this is a weak basis for the claim of real-world robustness and should be described as a preliminary result rather than a definitive validation.","section":"Table 6 / Section 4.3"},{"comment":"The comparison with NU-NeRF on TransLab (Table 9) reports that NU-NeRF collapses to a spherical shape; please provide the qualitative evidence in the appendix or state the convergence criterion used, since a single failure mode does not by itself establish method superiority.","section":"Appendix A.6"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely acceptable for a conference venue such as ACM MM if the overclaimed error bound is removed, the hyperparameters are reported, and the ablation discussion is made consistent with Table 4. For a journal version, the authors should also add variance reporting and a more balanced treatment of the synthetic benchmark versus the real-world validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a credible engineering contribution to transparent object reconstruction with 3DGS: a two-stage training scheme (geometry on de-lighted images, then appearance with frozen opacity) plus a max-weight sliding-window depth extraction at inference. Second, the paper's headline — the 37.3% CD reduction, the 8% F1 gain, and especially the 3 mm error bound — is oversold. The main benchmark is synthetic, disclosed only in the appendix; the key hyperparameters are unreported; and the ablation suggests the window-based depth extraction is not what carries the improvement.\n\nWhat's actually new: the sliding-window first-surface depth is a genuinely new heuristic, and the opacity-freezing split is a sensible combination. The TransLab dataset, though synthetic, is useful: 8 lab scenes with GT meshes, depth, normals, and masks. DTU results show the geometry stage generalizes to opaque objects, and the ClearPose appendix gives some real-world evidence, even if only one scene per set. Training at 0.8 hours and 105 FPS is good. I also give credit for the Table 8 comparison against baselines augmented with the same normal priors, which addresses an obvious confound.\n\nSoft spots, in order of seriousness. (1) The Appendix A.3 claim that depth error is naturally bounded by the window size, at most 3 mm, does not follow. The window maximizes the sum of T_i α_i inside a transmittance interval [T_end, T_start]; nothing ties that maximum to the true first surface when the foreground is highly transparent and the background is opaque. The bound holds only if the true surface falls inside W*, which is exactly the unproven assumption. (2) The mechanism is weakly supported by the ablation: replacing the first-surface extraction with PGSR's unbiased depth changes CD from 1.85 to 1.89 and leaves F1 unchanged. So the normal and de-light priors are doing the heavy lifting, and the paper should say so. (3) The TransLab dataset is synthetic but the main text says 'collect' without flagging it; a reader could reasonably assume real captures. That disclosure belongs in the abstract or intro. (4) Missing comparisons with TranSplat, Car-GS, alpha-NeuS, and alpha-surf — all cited in the related work, none in the tables. (5) No seed variance or error bars, and T_start, T_end, θ_T, θ_n are unreported.\n\nNone of this is fatal. The method is coherent, the code and dataset are promised, and the limitations section honestly admits the single-layer transparency assumption. The paper deserves serious peer review, but it needs revision: report all hyperparameters and seeds, add the missing baselines, disclose the synthetic nature prominently, and rewrite the 3 mm bound as a heuristic property rather than a theorem.\n\nFor whom: people working on transparent object reconstruction or robot manipulation of glassware will get value from the dataset and the two-stage recipe. Whether I'd cite it depends on the code and dataset actually shipping; the promises are good, but they need to be honored.","headline":"Solid engineering paper with a useful synthetic dataset, but the headline 3 mm error bound and the central depth-extraction mechanism are both weaker than claimed.","tokens_in":23392,"tokens_out":2896,"would_cite":true,"duration_ms":27938,"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":"Transparent surfaces get accurate 3D geometry by learning geometry separately from appearance and extracting first-surface depth with a sliding-window rule.","keywords":["3D Gaussian splatting","transparent surface reconstruction","first-surface depth extraction","accumulated transmittance","alpha-blending depth","normal and de-lighting priors","anisotropic spherical Gaussians","TransLab dataset"],"falsifier":"Construct a controlled scene with a flat transparent slab of known position, varying surface opacity and background texture contrast, and compare TSGS's extracted first-surface depth against the known slab distance: if depth error exceeds the 3 mm window size whenever the slab's alpha contribution is low, the window-containment assumption is violated. A simpler version is a filled beaker with an immersed object, where multi-layer transparency is present and the single-layer assumption predicts a biased first-surface depth.","tokens_in":22242,"feed_emoji":"🧪","tokens_out":7414,"duration_ms":71925,"temperature":0.7,"pith_summary":"This paper claims that the transparency-depth dilemma in 3D Gaussian Splatting—where alpha-blending weights tuned for realistic appearance push depth estimates of transparent objects behind the true surface—can be resolved without abandoning the 3DGS framework. It introduces TSGS, which learns geometry from de-lighted images and normal priors in a first stage, then refines appearance with anisotropic specular modeling while freezing opacity in a second stage. For depth, it replaces full alpha-blended depth with a sliding-window maximum-weight search over accumulated transmittance, extracting the first surface along each ray. On the new TransLab dataset of laboratory glassware, TSGS reports a 37.3% lower chamfer distance and an 8.0% higher F1 score than the strongest baseline, with a 0.41 dB PSNR gain, showing geometry and photorealistic transparency can be recovered together. This matters because robotic lab manipulation needs millimeter-accurate positions of transparent vessels.","feed_headline":"First-surface trick cuts 3D glassware errors by 37%","feed_subtitle":"A sliding window over blending weights finds the true surface, fixing depth while keeping real-time rendering.","key_machinery":"The load-bearing mechanism is the maximum-weight sliding-window first-surface extractor. For each pixel ray, after sorting intersecting Gaussians by depth, it computes accumulated transmittance $T_i$ and blending weight $T_i\\alpha_i$, keeps only the segment with $T_i$ between $T_{\\text{end}}$ and $T_{\\text{start}}$, slides a window of fixed size $\\delta_t$ (3 mm) over candidate Gaussians, and selects $W^* = \\arg\\max_j \\sum_{i\\in W_j} T_i\\alpha_i$. The first-surface depth is $D_{\\text{first}}(\\mathbf{u}) = (\\sum_{i\\in W^*} T_i\\alpha_i \\hat{d}_i(\\mathbf{u})) / (\\sum_{i\\in W^*} T_i\\alpha_i)$, using per-Gaussian plane depths rather than center depths. This is what isolates the first surface from background transmission and averages out floaters; the two-stage training, geometry on de-lighted images with normal priors followed by appearance refinement with opacity frozen and anisotropic spherical Gaussians, supplies the opacity field that makes the window meaningful.","core_discovery":"On its own terms, the discovery is that the opacity field learned purely for appearance still contains a reliable first-surface signal: the accumulated transmittance $T_i$ drops where Gaussians representing the first surface contribute, and the sum of $T_i\\alpha_i$ spikes there. TSGS locates this by restricting attention to the ray segment where $T_i$ lies between thresholds, sliding a fixed-size window along the sorted Gaussians, and picking the window with the largest total $T_i\\alpha_i$ weight. Depth is then the weighted average of per-Gaussian plane depths $\\hat{d}_i(\\mathbf{u}) = d_i/(\\mathbf{n}_i \\cdot \\mathbf{v}_\\mathbf{u})$ inside that window. Weighting by $T_i\\alpha_i$ within the window suppresses floaters that corrupt nearest-depth methods, while restricting to the window excludes background Gaussians seen through transparency that corrupt standard $\\alpha$-blended depth. The window size doubles as an error bound, set to 3 mm in the paper, so the reported accuracy figures rest on the window's ability to contain the true surface.","pith_inferences":["Editor's inference: The sliding-window rule could be applied as a drop-in depth estimator to other volume-rendered transparent-object reconstructions, since it only needs accumulated-transmittance weights that those renderers already compute.","Editor's inference: The stated 3 mm error bound is conditional on the maximum-weight window containing the true first surface; a stress test crossing single-layer versus nested transparent objects, such as liquid inside a beaker, would show how often that containment fails.","Editor's inference: Because the ablation shows the normal prior is the largest single geometry contributor, the method's accuracy is likely sensitive to the quality of that diffusion-derived prior; testing on scenes with heavy occlusion or unusual glassware would quantify this sensitivity.","Editor's inference: A direct extension would be multi-window extraction to recover the second surface, turning TSGS from a single-shell reconstruction into a layered refractive reconstruction; the paper's own limitation section points at this direction."],"forward_implications":["Transparent object geometry becomes extractable from standard appearance-optimized Gaussians at inference time, without a separate depth network or ray-tracing pass.","With the 3 mm window bound, the resulting depth error is controlled enough for millimeter-precision robotic manipulation of laboratory glassware, on the paper's reported results.","Freezing opacity after geometry learning prevents appearance refinement from eroding shape accuracy, so visual fidelity and geometry are not forced to trade off within a single optimization.","On TransLab, the full pipeline yields CD 1.85 and F1 0.95, compared with 2.95 and 0.88 for the strongest baseline, while keeping 105 FPS rendering and a training time near 0.8 hours.","The same geometry stage transfers to opaque objects: on DTU, TSGS reports a mean chamfer distance of 0.51, competitive with or better than the compared surface reconstruction methods."],"supporting_citations":[{"why":"Provides the base 3D Gaussian Splatting representation, alpha-blending rendering, and optimization loop that TSGS modifies.","marker":"[31]"},{"why":"Supplies the planar flatten regularization, Gaussian plane-depth formulation, and the unbiased-depth baseline that TSGS's first-surface extraction replaces.","marker":"[9]"},{"why":"Supplies the normal and de-light priors used to supervise geometry in Stage 1; removing them is the paper's largest ablation hit.","marker":"[75]"},{"why":"Provides the anisotropic spherical Gaussian appearance model used in Stage 2 to render specular and transmission effects while opacity is frozen.","marker":"[73]"},{"why":"Defines the 2D Gaussian splatting surface-reconstruction baseline compared on TransLab and DTU.","marker":"[25]"},{"why":"Gaussian opacity fields, a surface-reconstruction baseline compared on TransLab and DTU.","marker":"[79]"},{"why":"TSDF fusion routine used to turn rendered first-surface depth maps into meshes for evaluation.","marker":"[53]"},{"why":"DTU dataset on which TSGS demonstrates competitive geometry on opaque objects and establishes generalizability.","marker":"[27]"}],"fun_headline_variants":["Transparent surfaces: 37% better depth from a sliding window","Sliding-window depth trick fixes transparent 3D scans","TSGS: one trick rescues transparent surface geometry","37% less error in 3D glassware via first-surface depth","Sliding alpha window yields crisp depth for transparent objects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ray segment where the sliding-window sum of accumulated-transmittance-weighted alpha values is largest actually contains the true first surface; if a weak first-surface signal relative to transmitted background or poor threshold choices breaks that containment, the extracted depth is biased.","fun_headline_variants_meta":{"raw":{"variants":["Transparent surfaces: 37% better depth from a sliding window","Sliding-window depth trick fixes transparent 3D scans","TSGS: one trick rescues transparent surface geometry","37% less error in 3D glassware via first-surface depth","Sliding alpha window yields crisp depth for transparent objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000309,"raw_usage":{"total_tokens":1819,"prompt_tokens":1058,"completion_tokens":761,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":674,"completion_tokens_details":{"reasoning_tokens":676}},"tokens_in":674,"tokens_out":761,"duration_ms":7435,"temperature":1.0,"reasoning_tokens":676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:22:01.019950+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a controlled scene with a flat transparent slab of known position, varying surface opacity and background texture contrast, and compare TSGS's extracted first-surface depth against the known slab distance: if depth error exceeds the 3 mm window size whenever the slab's alpha contribution is low, the window-containment assumption is violated. A simpler version is a filled beaker with an immersed object, where multi-layer transparency is present and the single-layer assumption predicts a biased first-surface depth.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the normal and de-light priors used to supervise geometry in Stage 1; removing them is the paper's largest ablation hit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DTU dataset on which TSGS demonstrates competitive geometry on opaque objects and establishes generalizability."}],"review_version":1}