{"id":"a6bac246-3f3e-4047-9435-a858274b7bb2","arxiv_id":"2501.11508","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"SIDGaussian adds DINO-based semantic consistency and local Pearson-correlation depth regularization to sparse-view 3DGS, reaching 20.71 dB PSNR on LLFF, 0.4 dB above FSGS.","lead":"This paper proposes two extra training rules for 3D Gaussian Splatting when only a few photos of a scene are available, one that matches semantic features between training and novel views and one that aligns locally normalized depth patches. On the LLFF benchmark with three training views, the method reports small PSNR gains and larger perceptual gains over the FSGS baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (5) subtracts a minimized L1 norm of a Pearson correlation, which pushes rendered depth anti-correlated with DPT depth, contradicting the stated geometric-regularization intent.","rationale":"The reader's central objection is correct: the paper defines a local depth regularizer as the L1 norm of the Pearson correlation. Minimizing that L1 norm drives the correlation to zero, not toward +1, so the loss cannot implement the stated goal of making rendered depth similar to DPT depth. This is not a disagreement with a community consensus; it is an internal inconsistency between Eq. (5) and the surrounding text. The local-depth term is one of the two core contributions and is specifically credited in the ablation for improved geometric detail, so the paper as written does not support its central claim. A corrected loss (1 - Corr) might salvage the method, but that would be a change to the stated method. The reader's additional points about missing code, error bars, and the incorrect DINO reference are secondary but reinforce the moderate-confidence rejection.","tokens_in":7831,"tokens_out":1427,"duration_ms":14994,"concrete_test":"Re-derive the gradient of Eq. (5) with respect to the rendered depth patch dhat_LN, or implement the term in PyTorch on a small patch: initialize dhat_LN = dLN, evaluate Ldepth and its gradient, and take one gradient step. With the paper's literal ||Corr||_1 formulation, the step should leave Corr at 0 and push patch variance toward 0. If instead the paper meant 1 - Corr (or -Corr), the step should increase Corr toward +1. The correct version will make the intended geometric prior nonzero; the literal version will not.","verdict_should_be":"REJECT","load_bearing_attack":"The central geometric regularization is Eq. (5): Ldepth = ||Corr(dLN, dhat_LN)||_1, presented as encouraging similarity between locally normalized DPT depth dLN and rendered depth dhat_LN. Because Corr is the Pearson correlation, and ||·||_1 of a scalar is its absolute value, minimizing this loss drives the correlation toward zero, not toward +1. A correlation of 0 means the rendered local depth is the best affine approximation of the DPT depth, but is otherwise decoupled in shape; there is no gradient favoring positively correlated structure. The text and Fig. 2 state that the method 'encourages the similarity between local patches' and uses correlation 'to measure the similarity,' which would require penalizing 1 - Corr or -Corr. As written, the loss cannot implement the described prior; its main gradient would push rendered depth toward locally constant patches (zero variance), actively destroying geometric detail rather than preserving it. The ablation in Table II is consistent with a method that still benefits from DINO semantic regularization or from other effects, but it cannot demonstrate the local-depth mechanism claimed, because the stated term does not do what the claim requires.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents SIDGaussian, a sparse-view 3D Gaussian Splatting method that augments the FSGS baseline with two regularizers. A semantic regularization term (Eq. (3)) penalizes the L2 distance between DINO-ViT features of patches from rendered training and side views, and a local depth regularization term (Eq. (5)) is defined as the L1 norm of the Pearson correlation between locally normalized DPT depth patches and rendered depth patches. Experiments on LLFF with three training views report state-of-the-art PSNR/SSIM/LPIPS at 1/8 and 1/4 resolution, with an ablation in Table II showing gains from both semantic and depth terms. The paper claims up to 0.4 dB PSNR improvement over FSGS.","tokens_in":8015,"tokens_out":8055,"duration_ms":85230,"significance":"If the contributions were correctly formulated, SIDGaussian would be a modest but useful step for real-time sparse-view rendering: the semantic prior is principled, the comparison is direct, and held-out-view evaluation with external baselines avoids circularity. The reported numerical gains are positive across all three metrics. However, the principal geometric prior is written incorrectly, and the DINO-ViT citation is wrong; both must be corrected before the empirical claims can be interpreted. With the current Eq. (5), the depth term does not implement the described similarity objective, so the ablation's depth row cannot be credited to the stated mechanism. The paper is not acceptable in this form.","major_comments":[{"comment":"Equation (5) defines L_depth = ||Corr(d_LN, d_hat_LN)||_1, with Corr the Pearson correlation. Since Corr lies in [-1,1], minimizing this absolute value drives the correlation toward 0, not toward +1. The text states that the loss \"encourages the similarity between local patches\" and uses correlation \"to measure the similarity\"; that objective would require minimizing 1 - Corr or -Corr. Under the equation as written, the gradient prefers zero correlation (or a zero-variance rendered patch), which is the opposite of aligning the rendered depth structure with the DPT prediction. Consequently, the Table II ablation row labeled \"local depth regularization\" does not demonstrate the claimed mechanism. This is the central load-bearing issue for the depth contribution.","section":"II-B2, Eq. (5)"},{"comment":"The semantic regularizer is claimed to use \"DINO-ViT [40]\", but reference [40] is \"DINO: Detr with improved denoising anchor boxes for end-to-end object detection\" (arXiv:2203.03605), not the self-supervised DINO ViT feature extractor of Caron et al. The paper must cite the correct pretrained model and specify which layer and patchification are used; as written, the semantic regularizer is not reproducible from the bibliography.","section":"II-B1, reference [40]"},{"comment":"The depth prior assumes DPT predictions for rendered side views are reliable. These side views come from an under-constrained 3DGS model that is itself being optimized, and DPT is not trained on synthetic rendered images. The paper provides no validation of this assumption, e.g., by comparing DPT side-view depth against ground-truth depth on held-out views or by ablating the choice of side-view source. Even after Eq. (5) is corrected, this correctness risk remains and should be analyzed.","section":"II-B2 and III-A"}],"minor_comments":[{"comment":"The loss weights omega_0, omega_sem, and omega_depth are not reported; only the patch size and iteration count are given. Specify all hyperparameters for reproducibility.","section":"III-A"},{"comment":"The function d(.) in Eq. (2) is called a \"distance function\" but is never defined, while d_LN and d_hat_LN are later used for depth maps; the notation should be disambiguated.","section":"II-A, Eq. (2)"},{"comment":"The symbol P is first used as a set of patches (P ⊆ {Pr, Pt}) and then as the spatial support of x in the normalization formula; this overloaded notation should be clarified.","section":"II-B2, Eq. (4)"},{"comment":"References [43] and [48] are the same paper and should be merged; the abstract also contains the typo \"inphut\".","section":"References"},{"comment":"The paper claims real-time rendering but reports no runtime; Fig. 5 also does not specify the axes or the exact weight values used, so the sensitivity discussion is difficult to reproduce.","section":"III-B and Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Eq. (5) is valid and is reflected in major comment 1. I do not think outright rejection is necessary, because the framework and semantic prior could be salvaged with a corrected depth loss and rerun experiments; however, no part of the current depth mechanism can be approved as written, and the incorrect DINO citation further weakens reproducibility. The duplicate reference [43]/[48] also suggests the bibliography was not checked carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2501.11508 (SIDGaussian): the paper proposes two regularizers for sparse-view 3D Gaussian Splatting—DINO-based semantic consistency and a local depth correlation term. The semantic piece is sensible, and the LLFF results look decent (20.71 PSNR at 1/8 res, beating FSGS by 0.4 dB). The problem is that Eq. (5) is written as the L1 norm of a Pearson correlation. Minimizing that drives the correlation to zero, not toward +1. So the stated loss does not encourage similarity between rendered and DPT depth patches; it does the opposite. This is not a nitpick—it is one of the two claimed contributions. The ablation shows a 0.14 dB gain from this term, but the mechanism described in the text and Figure 2 cannot be what is happening. Either the equation is a typo (should be 1−Corr or −Corr) or the method is not what it claims. As written, the paper does not support its central claim.\n\nWhat is genuinely new: combining DINO semantic features with local depth normalization inside 3DGS training for sparse views is a reasonable idea, and the empirical comparison is a real result. The paper is clearly written and honest about the modest gains. Credit for that.\n\nSoft spots, in proportion: the depth-loss sign issue is load-bearing. Also missing code, no error bars, and a mistaken reference [40]—they cite the DETR-based DINO, not the self-supervised DINO-ViT they actually use. Those are minor next to the equation problem.\n\nThe stress-test note is correct on the equation. The paper may be fixable by changing Eq. (5), but it needs major revision before anyone can trust the local-depth claim. I don't agree with a clean reject purely on novelty grounds; the idea is worth a referee's time.\n\nWorth a quick look if you work on sparse-view 3DGS, mostly as a cautionary tale about sign errors. I wouldn't cite it in its current form. A serious editor should still send it to review—the core idea is salvageable and the subfield is active.","headline":"Equation (5) makes the local depth regularizer push correlation to zero instead of +1, contradicting the paper's own description and undermining one of its two central contributions; the semantic regularization is reasonable and the paper is salvageable, but not in current form.","tokens_in":8612,"tokens_out":2595,"would_cite":false,"duration_ms":26367,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that sparse-view 3D Gaussian Splatting can be substantially improved by adding a DINO-ViT semantic-consistency loss and a DPT-based local depth loss, outperforming prior methods by up to 0.4 dB PSNR on LLFF.","keywords":["Sparse-view novel view synthesis","3D Gaussian Splatting","Semantic regularization","Local depth regularization","DINO-ViT features","DPT depth priors","LLFF dataset","Multi-view consistency"],"falsifier":"Render side views from the trained model, compute the local Pearson correlation between the model's rendered depth and DPT's depth on those same views, and check whether the scenes with the largest PSNR gains also have the highest correlations; if improved scenes show near-zero correlation, the depth term is not doing the geometric work claimed.","tokens_in":7593,"feed_emoji":"🖼️","tokens_out":10728,"duration_ms":100284,"temperature":0.7,"pith_summary":"Sparse-view 3D Gaussian Splatting (3DGS) produces distorted geometry and lost detail when only a handful of images are available. This paper proposes a method, called SIDGaussian, that adds two prior-based losses to the standard sparse-view 3DGS training loop: a semantic consistency loss that aligns DINO-ViT features between rendered training views and rendered side views, and a local depth loss that matches rendered side-view depth patches to DPT monocular depth estimates in a scale-invariant way. The aim is to make sparse-input 3DGS competitive with NeRF-based sparse-view methods while retaining real-time rendering. On the LLFF dataset with three training views, the reported result is a PSNR of 20.71 dB at 1/8 resolution, a gain of 0.4 dB over the strongest prior method, together with the lowest LPIPS among all compared methods.","feed_headline":"Sparse-view 3D Gaussian Splatting gains 0.4 dB on LLFF with two priors","feed_subtitle":"Local depth patches and DINO semantic features keep three-view rendering consistent and detailed.","key_machinery":"Two auxiliary loss terms grafted onto the FSGS-style objective. The semantic regularizer uses a pretrained DINO-ViT encoder as a fixed perceptual teacher: it embeds randomly cropped patches from rendered training views and rendered side views and minimizes their L2 distance, so side views are pulled toward the semantic content of the training views. The depth regularizer uses a pretrained Dense Prediction Transformer (DPT) as a monocular depth teacher; it locally normalizes rendered and predicted depth patches with $d_{\\mathrm{LN}}(x) = (d(x)-\\mu)/(\\sigma+\\epsilon)$ and measures their similarity with Pearson correlation, making the geometric constraint invariant to depth scale and offset while still sensitive to local shape. The two losses are combined with the base color, SSIM, and global depth loss in a weighted sum, so the regularizers act directly on rendered images and require no extra data at inference time.","core_discovery":"The central claim is that an under-constrained 3DGS scene can be stabilized by supervising two things the standard loss never sees: semantic agreement across views and local relative depth structure. The semantic loss $L_{\\mathrm{sem}}$ computes the L2 distance between DINO-ViT embeddings of randomly cropped patches from training-view and side-view renderings, encouraging unseen views to carry the same content as the observed views. The local depth loss $L_{\\mathrm{depth}}$ renders depth for side views, normalizes each local 126x126 patch by its own mean and standard deviation, and uses Pearson correlation to compare the normalized patches with DPT predictions. Because both patch normalization and Pearson correlation remove scale and shift, the depth term constrains relative structure rather than absolute metric depth. In Table I the full method reports PSNR 20.71 / SSIM 0.708 / LPIPS 0.205 at 1/8 resolution, and the ablation in Table II shows semantic regularization alone lifts PSNR from 20.31 to 20.57 while the addition of local depth reaches 20.71.","pith_inferences":["The training loop is effectively a self-distillation scheme in which DINO and DPT act as fixed teachers for views the model has never seen; a natural stress test is two-view or single-view input, where the priors are weaker but matter more.","Because the depth term is invariant to per-patch scale and shift, it cannot correct global depth drift or metric scale errors; pairing it with an ordinal or ranking-based depth loss could close that gap.","The gains are reported only on LLFF's forward-facing scenes; unbounded or object-centric captures would test whether the semantic and depth priors survive large scale variation and occlusions.","The semantic loss's reliance on random patch crops suggests a simple extension: multi-scale crops or attention-weighted feature maps from several DINO layers could strengthen the consistency signal."],"forward_implications":["Three-view 3DGS becomes a more viable real-time alternative to sparse-view NeRF methods on forward-facing scenes, since the extra losses add no inference-time cost.","The semantic term alone accounts for most of the perceptual improvement in the ablation (LPIPS drops from 0.288 to 0.220), suggesting that enforcing semantic consistency across viewpoints is a strong prior for ill-posed radiance-field optimization.","Local depth regularization adds a smaller but consistent gain on top of the semantic term (PSNR from 20.57 to 20.71), indicating that scale-invariant depth matching captures detail that global depth supervision misses.","The total loss remains a simple weighted sum, so each regularizer can be switched on or off per scene while the base 3DGS pipeline stays unchanged."],"supporting_citations":[{"why":"Base sparse-view 3DGS method this work extends; supplies the side-view sampling strategy and the global depth loss term.","marker":"[39]"},{"why":"Supplies the pretrained DPT monocular depth estimator whose locally normalized patches act as geometric supervision for the depth regularization.","marker":"[37]"},{"why":"Cited as the source of the DINO-ViT encoder used to compute semantic features for the semantic regularization.","marker":"[40]"},{"why":"Defines Pearson correlation, the scale-invariant similarity metric used to compare rendered and DPT depth patches.","marker":"[42]"},{"why":"Provides the LLFF dataset and the image-split convention used for all reported quantitative comparisons.","marker":"[49]"},{"why":"Establishes the three-view evaluation protocol and serves as a NeRF-based baseline with depth smoothness regularization.","marker":"[33]"},{"why":"Baseline that distills DPT depth rankings into NeRF; the paper reports a 0.95 dB gain over it.","marker":"[36]"},{"why":"Few-shot NeRF baseline included in the quantitative and visual comparisons.","marker":"[38]"}],"fun_headline_variants":["Sparse-view 3DGS gets 0.4 dB boost with dual priors","Local depth and DINO semantics fix sparse-view 3DGS","Under-constrained 3DGS stabilized by semantic and depth priors","Sparse-view rendering improved by semantic and depth regularization","3DGS sparse view: 0.4 dB gain from DINO and local depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that DPT monocular depth for the rendered side views is accurate enough to supervise geometry even though the 3DGS model generating those side views is still being trained; if DPT is systematically biased on those synthetic views, the local depth loss would reinforce wrong geometry rather than correct it.","fun_headline_variants_meta":{"raw":{"variants":["Sparse-view 3DGS gets 0.4 dB boost with dual priors","Local depth and DINO semantics fix sparse-view 3DGS","Under-constrained 3DGS stabilized by semantic and depth priors","Sparse-view rendering improved by semantic and depth regularization","3DGS sparse view: 0.4 dB gain from DINO and local depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":2982,"prompt_tokens":929,"completion_tokens":2053,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":1954}},"tokens_in":545,"tokens_out":2053,"duration_ms":16518,"temperature":1.0,"reasoning_tokens":1954,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:10:03.834781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render side views from the trained model, compute the local Pearson correlation between the model's rendered depth and DPT's depth on those same views, and check whether the scenes with the largest PSNR gains also have the highest correlations; if improved scenes show near-zero correlation, the depth term is not doing the geometric work claimed.","supporting_citations":[{"cited_title":"Fsgs: Real-time few-shot view synthesis using gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Base sparse-view 3DGS method this work extends; supplies the side-view sampling strategy and the global depth loss term."},{"cited_title":"Vision transformers for dense prediction,","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained DPT monocular depth estimator whose locally normalized patches act as geometric supervision for the depth regularization."},{"cited_title":"Local light field fusion: Practical view synthesis with prescriptive sampling guidelines,","cited_arxiv_id":null,"evidence_quote":"Provides the LLFF dataset and the image-split convention used for all reported quantitative comparisons."},{"cited_title":"Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs,","cited_arxiv_id":null,"evidence_quote":"Establishes the three-view evaluation protocol and serves as a NeRF-based baseline with depth smoothness regularization."},{"cited_title":"Sparsenerf: Distilling depth ranking for few-shot novel view synthesis,","cited_arxiv_id":null,"evidence_quote":"Baseline that distills DPT depth rankings into NeRF; the paper reports a 0.95 dB gain over it."}],"review_version":1}