{"id":"d925ab79-f9b6-4e92-b41e-f7e222c67403","arxiv_id":"2508.04099","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"DET-GS combines hierarchical depth normalization, Canny-guided depth smoothing, and RGB-gated total variation to modestly improve 3DGS rendering quality on five standard datasets.","lead":"DET-GS adds three regularization losses to 3D Gaussian Splatting: hierarchical depth supervision, edge-masked depth smoothing, and an RGB-guided total variation term. On standard full-view benchmarks the method posts small but consistent PSNR/SSIM/LPIPS gains, but the abstract's sparse-view claim is not actually tested.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed sparse-view SOTA gains are untested: all experiments use full training sets, so the abstract's central claim lacks supporting evidence.","rationale":"I read the paper in good faith: it introduces three plausible regularizers and provides full-set experiments, ablations, and per-scene tables. However, the central claim in the abstract is specifically about sparse-view benchmarks, and the experiments section contains no sparse-view evaluation whatsoever. This mismatch between claim and evidence is the most load-bearing concern because it directly determines whether the paper delivers on its advertised contribution. The reader's rationale explicitly flags this missing sparse-view experiment, but the reader's stated 'weakest assumption' focuses on monocular depth reliability; hence I partially agree. A single focused sparse-view experiment would settle the issue: if the method does not help under sparse views, the paper's main promise fails; if it does, the rejection could be revised to conditional acceptance pending the other under-specification fixes. Since the current evidence does not support the abstract's claim, I uphold the REJECT verdict without changing it.","tokens_in":17249,"tokens_out":4565,"duration_ms":53733,"concrete_test":"Re-run the Mip-NeRF 360 experiments on a sparse-view protocol, e.g., 8 training views per scene (as in SparseGS/FSGS), keeping all other training settings identical (30k iterations, same baselines). Compare PSNR/SSIM/LPIPS against 3DGS, DNGaussian, and Spec-Gaussian under the same subset. If DET-GS does not outperform these baselines by margins comparable to Table I, the abstract's sparse-view claim is unsupported. Also run the LLFF 3-view/6-view protocol from DNGaussian to confirm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and introduction repeatedly identify sparse-view novel view synthesis as the core problem, and the strongest claim is that DET-GS 'outperforms state-of-the-art (SOTA) methods on sparse-view novel view synthesis benchmarks.' However, Section IV describes experiments only on the full training sets of Mip-NeRF 360, Tanks&Temples, Deep Blending, NeRF Synthetic, and NSVF. No experiment subsamples training views (e.g., 3, 6, or 9 views), and no standard few-shot benchmark such as LLFF, DTU, or the sparse-view protocols of DNGaussian/FSGS is used. Consequently, the headline claim is not supported by any evidence in the paper. The method may improve dense reconstruction, but the paper's stated purpose is sparse-view improvement, and that remains unverified. This is load-bearing because if DET-GS fails to outperform prior SOTA on an actual sparse-view protocol, the central contribution collapses. Secondary under-specification issues (e.g., the undefined omega in Eq. 4 and the undeclared 'error tolerance mechanism') compound the problem but are not the focus here.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DET-GS, a 3D Gaussian Splatting extension with three regularization losses: (i) hierarchical depth supervision using patch-wise and image-normalized L2 losses against Depth Anything V2 predictions (Eqs. 4-7); (ii) edge-aware depth smoothing masked by Canny edges (Eqs. 8-11); and (iii) RGB-guided edge-preserving total variation (Eqs. 13-15). These are combined with a color loss in Eq. (17). Experiments compare on Mip-NeRF 360, Tanks&Temples, Deep Blending, NeRF Synthetic, and NSVF, plus ablations. The abstract claims sparse-view SOTA and geometric accuracy gains.","tokens_in":17499,"tokens_out":8735,"duration_ms":107010,"significance":"The paper addresses a relevant problem: using monocular depth priors to regularize 3DGS, and the three components are clearly motivated. The clean decomposition of the objective and the use of an external depth prior avoid circularity. If the sparse-view claim were demonstrated, the method would be a useful contribution to 3DGS regularization. However, as submitted the significance is conditional: no code is provided, geometric accuracy is not measured quantitatively, and the core sparse-view claim is not tested. The numerical gains on dense-view benchmarks are modest (e.g., +0.17 dB PSNR on Mip-NeRF360 vs Spec-Gaussian), which is insufficient by itself to support the stated contribution.","major_comments":[{"comment":"The abstract and introduction claim 'outperforming SOTA on sparse-view novel view synthesis benchmarks,' but all experiments use the full training sets of Mip-NeRF 360, Tanks&Temples, Deep Blending, NeRF Synthetic, and NSVF. There is no view subsampling (e.g., 3/6/9 views) and no sparse-view protocol such as LLFF/DTU or the DNGaussian/FSGS few-shot settings. Thus the central claim is not supported by any experiment in the paper. Please either add standard sparse-view experiments or revise the claims to dense-view benchmarks.","section":"Section IV.C (Tables I-III)"},{"comment":"The paper claims improvements in 'geometric accuracy' in the abstract and conclusion, but the only quantitative metrics are PSNR/SSIM/LPIPS. Depth maps are compared only qualitatively in Fig. 6. No metric such as absolute relative depth error, Chamfer distance, F-score, or comparison against a SfM/COLMAP geometry baseline is reported. Consequently the geometric-accuracy claim is not substantiated.","section":"Section IV.C"},{"comment":"The modified depth rendering in Eq. (4) uses an undefined scalar ω and replaces the alpha-composited weights of Eq. (3) with ω exp(−ι(n−1)). There is no statement of the value, range, or schedule of ω; the phrase 'enhanced opacity values ω to all primitives' is difficult to interpret because G2D_n is already part of the projection, while the per-Gaussian opacity α_n is omitted. Since this depth is used in L_depth and L_edge, the failure to specify and justify this weighting is load-bearing. Please align with Eq. (3) or define and validate the new weighting.","section":"Section III.B, Eq. (4)"},{"comment":"The implementation section states: 'To mitigate over-constraining the learning process, we introduce an error tolerance mechanism into the depth L2 loss,' but no equation, threshold, or algorithmic description is given in Section III or Eq. (17). The loss in Eq. (7) is a plain L2 loss. This unspecified mechanism can affect all quantitative results and the ablation; without specification, the experiments are not reproducible.","section":"Section IV.B, 'error tolerance mechanism'"}],"minor_comments":[{"comment":"'where µ_i denotes the center position' should be µ_n; also the relationship between Eq. (3) and Eq. (4) is confusing because both are called D(x).","section":"Section III.B, Eq. (4)"},{"comment":"The phrase 'semantic masks derived from Canny edge detection' is inaccurate; Canny outputs an edge map, not semantic masks. Please reword.","section":"Abstract / Introduction"},{"comment":"It is unclear what 'patch size 0' means and how 'randomized patch size sampling' is compatible with 'non-overlapping patches.' Clarify the protocol.","section":"Section IV.D, Table V"},{"comment":"Scaffold-GS has PSNR 30.20 versus Ours 30.05, so the text 'consistently outperforms previous SOTA methods across all datasets' should be qualified (e.g., best on average, or per metric).","section":"Table I, Deep Blending column"},{"comment":"The labels and numbers in the figure are hard to read; clarify which variant corresponds to each PSNR/LPIPS value and why the best scores are marked in red.","section":"Fig. 5"},{"comment":"When all neighbors in the cross-shaped neighborhood are masked (m=0), the denominator reduces to epsilon and the masked mean becomes 0. Consider including a validity count to avoid numerical instability.","section":"Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the mismatch between the paper's claims and its evidence. The requested sparse-view evaluation is a substantial addition, but it is necessary for the stated contribution. I would not recommend acceptance before that evidence is provided; if the authors cannot demonstrate gains under a standard few-shot protocol, the contribution reduces to dense-view regularization with modest improvements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent incremental 3DGS paper with a mismatch between its headline claim and its experiments. The method — hierarchical patch-wise depth normalization plus Canny-masked depth smoothing plus RGB-gated TV — is clearly explained, and the ablations show each piece earns its keep on standard benchmarks. But the abstract and intro promise sparse-view SOTA gains, and no sparse-view experiment exists in the paper. That is a load-bearing gap, not a style issue.\n\nWhat it does well: the losses are clean, and the ablation table is honest about contributions. The gains over Spec-Gaussian and others on Mip-NeRF 360, Tanks&Temples, NeRF Synthetic, and NSVF are consistent (on the order of 0.1–0.3 dB PSNR and a few hundredths of LPIPS). The depth rendering formula in Eq. 4 is a reasonable modification to emphasize the nearest Gaussians, though the weight ω is never given a value or a sweep.\n\nThe soft spots are real. First, no view subsampling anywhere in Section IV; all five datasets are trained on the full sets. So the central sparse-view claim is unsupported. That alone makes the paper unacceptable in current form. Second, the 'error tolerance mechanism' in the depth loss is mentioned but never specified — a reader cannot reproduce what exactly is being done. Third, the paper claims geometric accuracy but reports only RGB metrics; the depth maps are illustrative. Fourth, the related work criticizes non-local depth losses without acknowledging that DNGaussian, the most relevant baseline, already uses local-global depth normalization; the proposed hierarchical normalization is closer to a variant than a wholesale fix. None of these are fatal to the method itself, but they are exactly the things a referee would need resolved.\n\nThe method is a plausible dense-view regularization stack with small but consistent gains. The writing is clear, and the ablations are competently done. But the sparse-view framing is central and untested; a revised version that actually subsamples views (e.g., 3/6/9 on LLFF or DTU, or the DNGaussian protocol) and fixes the ω/error-tolerance spec would be worth a second look. As it stands, I'd send it to peer review — the issue is fixable and worthwhile — but the current submission should be rejected or major-revisioned.\n\nMy recommendation: treat as an interesting incremental contribution with a significant evidentiary gap. Not something I'd cite yet, but worth bringing to a reading group to discuss claims-evaluation in graphics papers.","headline":"Competent incremental 3DGS paper whose headline sparse-view claim is untested — the experiments all use full training sets, so the central motivation is unsupported.","tokens_in":18005,"tokens_out":2415,"would_cite":false,"duration_ms":27953,"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":"DET-GS, an extension of 3D Gaussian Splatting, claims that adding hierarchical depth supervision, Canny-edge-masked depth smoothing, and RGB-guided total variation improves sparse-view geometry and rendering quality.","keywords":["3D Gaussian Splatting","novel view synthesis","sparse-view reconstruction","depth supervision","monocular depth estimation","edge-aware regularization","total variation","Canny edge detection"],"falsifier":"Take the DET-GS pipeline and corrupt the depth prior, for example by adding noise only in texture-less regions or by applying a global scale shift, then measure PSNR/SSIM/LPIPS against the no-depth-supervision baseline; if the metrics do not drop, the claimed benefit is not coming from the depth prior. Separately, ablate the re-weighting by sweeping $\\omega$ from near 0 to near 1 on a fixed Mip-NeRF 360 scene: if performance is flat across the range, the exponential depth re-weighting in Eq. 4 is not load-bearing.","tokens_in":17095,"feed_emoji":"🖼️","tokens_out":9053,"duration_ms":95425,"temperature":0.7,"pith_summary":"The paper proposes DET-GS, an extension of 3D Gaussian Splatting that adds three regularization losses aimed at sparse-view novel view synthesis. The central claim is that these losses—hierarchical depth supervision against monocular depth estimates, edge-aware depth smoothing masked by Canny edges, and an RGB-guided edge-preserving total variation term—together improve both the recovered geometry and the rendered images, reducing floaters and preserving sharp boundaries. The authors report consistent improvements over several prior 3DGS-based methods on five benchmark datasets, and their ablation attributes the largest gain to the hierarchical depth supervision. If the claim holds, it means noisy monocular depth can serve as a reliable geometric prior when supervision is localized and error-tolerant, and that structure-aware smoothing can replace uniform smoothing in point-based rendering.","feed_headline":"Depth and edge losses lift sparse-view 3D Gaussian splatting","feed_subtitle":"Edge-masked depth and image-space regularization sharpen geometry and cut floaters.","key_machinery":"The carrying mechanism is a triple of auxiliary losses attached to the standard 3DGS optimization. The first is hierarchical depth supervision built on a re-weighted depth render (Eq. 4) that up-weights the nearest Gaussian through a parameter $\\omega$, with the rendered and estimated depths compared after patch-level and image-level normalization. The second is edge-aware depth regularization: a binary mask from Canny edge detection selects non-boundary pixels, and a masked local mean over a cross-shaped kernel smooths only those pixels. The third is RGB-guided edge-preserving total variation: binary masks derived from the ground-truth RGB gradients disable the TV penalty at strong edges. T","core_discovery":"The core claim is that geometry and appearance in 3D Gaussian Splatting can both be improved under sparse views by replacing non-local depth losses and uniform smoothing with structure-aware counterparts. Depth is rendered with a modified accumulation that gives the nearest Gaussian exponentially more weight, then compared with monocular depth predictions after patch-wise and image-wise normalization, creating a dual-scale supervision that drives Gaussian centers toward consistent positions. A second loss smooths rendered depth only where the Canny edge map is zero, using a masked local mean over a four-connected neighborhood. A third loss applies total variation to the rendered image only w","pith_inferences":["A natural extension is to replace the Canny edge mask with learned or confidence-weighted boundaries; the paper's own limitation section says the hand-crafted edge regularizer may not generalize across all scene types and lighting conditions.","The same patch/image-normalized supervision recipe could be applied to other geometry priors, such as surface normals or segmentation planes, in any point-based renderer.","The re-weighted depth render introduces a parameter $\\omega$ that the paper does not ablate; sweeping it would reveal whether the exponential reweighting or the hierarchical normalization carries the improvement.","Because all three losses act entirely in image and depth space, they should transfer to any differentiable rasterizer that outputs depth, suggesting the gains are not tied to the specific 3DGS implementation."],"forward_implications":["Sparse-view 3DGS reconstructions will show fewer floating artifacts and sharper depth discontinuities than methods that apply uniform smoothing.","The hierarchical depth term, not the smoothing terms, is the main driver of the reported gains, according to the paper's ablations.","RGB-guided total variation smooths flat regions while preserving texture, so rendered images should be perceptually cleaner in homogeneous areas.","Canny-based edge masking preserves object contours that uniform depth smoothing blurs.","The hierarchical depth term makes the optimization more robust to depth-estimation noise than pixel-to-pixel non-local depth losses."],"supporting_citations":[{"why":"Defines the base 3D Gaussian Splatting representation and optimization that DET-GS extends; supplies the color and depth rendering equations.","marker":"[7]"},{"why":"Supplies the monocular depth predictions used as pseudo-ground-truth for the hierarchical depth supervision loss.","marker":"[54]"},{"why":"Canny edge detector that produces the boundary masks guiding edge-aware depth regularization.","marker":"[15]"},{"why":"DNGaussian, a prior sparse-view 3DGS method using global-local depth normalization; the main baseline and motivation for local depth supervision.","marker":"[14]"},{"why":"Spec-Gaussian, a strong appearance-focused 3DGS baseline that the method must outperform across all datasets.","marker":"[10]"},{"why":"Mip-Splatting, an anti-aliasing 3DGS baseline used for comparison and source of some competing metrics.","marker":"[8]"},{"why":"Scaffold-GS, a structured 3DGS baseline used for comparison.","marker":"[9]"},{"why":"Mip-NeRF 360 dataset, one of the real-world benchmarks where sparse-view performance is measured.","marker":"[20]"},{"why":"NeRF Synthetic dataset, used for benchmark comparisons and qualitative results.","marker":"[1]"},{"why":"LPIPS perceptual metric used as an evaluation measure.","marker":"[63]"}],"fun_headline_variants":["Edge-aware depth and TV losses sharpen 3D Gaussian splatting","Depth and edge losses cut floaters in sparse-view 3DGS","Hierarchical depth and edge masks enhance 3D Gaussian splatting","Edge-masked depth and TV losses enhance 3DGS"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The depth-supervision branch assumes that the pretrained monocular depth predictor's maps are trustworthy in exactly the places that sparse-view reconstruction finds hardest, texture-less and occluded regions; the paper itself notes these maps can be inaccurate there, and Gaussian placement inherits those biases.","fun_headline_variants_meta":{"raw":{"variants":["Edge-aware depth and TV losses sharpen 3D Gaussian splatting","Depth and edge losses cut floaters in sparse-view 3DGS","Hierarchical depth and edge masks enhance 3D Gaussian splatting","Edge-masked depth and TV losses enhance 3DGS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":2846,"prompt_tokens":731,"completion_tokens":2115,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":2038}},"tokens_in":475,"tokens_out":2115,"duration_ms":17495,"temperature":1.0,"reasoning_tokens":2038,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:52:51.357565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the DET-GS pipeline and corrupt the depth prior, for example by adding noise only in texture-less regions or by applying a global scale shift, then measure PSNR/SSIM/LPIPS against the no-depth-supervision baseline; if the metrics do not drop, the claimed benefit is not coming from the depth prior. Separately, ablate the re-weighting by sweeping $\\omega$ from near 0 to near 1 on a fixed Mip-NeRF 360 scene: if performance is flat across the range, the exponential depth re-weighting in Eq. 4 is not load-bearing.","supporting_citations":[{"cited_title":"Depth anything v2,","cited_arxiv_id":null,"evidence_quote":"Supplies the monocular depth predictions used as pseudo-ground-truth for the hierarchical depth supervision loss."},{"cited_title":"A computational approach to edge detection,","cited_arxiv_id":null,"evidence_quote":"Canny edge detector that produces the boundary masks guiding edge-aware depth regularization."},{"cited_title":"Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization,","cited_arxiv_id":null,"evidence_quote":"DNGaussian, a prior sparse-view 3DGS method using global-local depth normalization; the main baseline and motivation for local depth supervision."},{"cited_title":"Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Spec-Gaussian, a strong appearance-focused 3DGS baseline that the method must outperform across all datasets."},{"cited_title":"Mip-splatting: Alias- free 3d gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Mip-Splatting, an anti-aliasing 3DGS baseline used for comparison and source of some competing metrics."}],"review_version":1}