{"id":"3f70fdfa-c82a-4df0-a05e-922224855255","arxiv_id":"2412.10051","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TSGaussian couples YOLOv9+SAM mask guidance with depth regularization to improve sparse-view 3D Gaussian Splatting reconstruction of specified target objects.","lead":"TSGaussian is a 3D reconstruction system that isolates one object and uses semantic masks plus monocular depth priors to rebuild it from sparse photos. It reports large gains over prior Gaussian Splatting baselines on target-specific novel view synthesis, a practical step toward phone-based 3D scanning of individual objects.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 contradicts the claimed all-metric superiority: on garden TSGaussian's SSIM/LPIPS are worse than 3DGS (0.942/0.049 vs 0.945/0.042), and on bouquet LPIPS is worse than Gaussian Grouping (0.128 vs 0.127).","rationale":"I read the central claim as empirical: TSGaussian outperforms state-of-the-art methods on sparse-view target-specific novel view synthesis. That requires the comparison table to be reliable. The most load-bearing point is not the accuracy of the monocular depth prior but the integrity of the quantitative comparison itself: the paper's own numbers contradict its own prose. The reader's depth-prior concern is legitimate but secondary, because the ablation shows depth regularization contributes only 0.496 PSNR on a single scene, so even an imperfect depth estimator is not the sole linchpin of the reported gains. The direct contradiction in Table 1 is closer to the empirical claim: if those numbers stand, 'outperforms all baselines in SSIM, LPIPS, and PSNR' is simply false. This could be a typo or an evaluation bug rather than misconduct, but it must be settled before the central claim can be accepted. The reader already assigned CONDITIONAL and flagged the table inconsistencies in the rationale, so my check strengthens that assessment rather than changing it. A single reproduction check would determine whether the concern lands: rerun all baselines and TSGaussian under identical splits, masks, resolution, and metrics, and compare per-scene values. Until then, the manuscript's strongest sentence should not be taken at face value.","tokens_in":9837,"tokens_out":8534,"duration_ms":92070,"concrete_test":"Reproduce Table 1 for all four scenes using the authors' exact train/test view splits and target masks, running the official implementations of 3DGS, DNGaussian, SparseGS, and Gaussian Grouping with identical rendering resolution and evaluation code. Specifically verify whether the garden 3DGS entry (26.05/0.945/0.042 vs TSGaussian 27.93/0.942/0.049) and the bouquet Gaussian Grouping LPIPS (0.127 vs 0.128) survive. If the corrected table preserves these counterexamples, the prose claim of universal all-metric superiority and of consistently lower LPIPS must be withdrawn, and the central claim should be recast as scene-dependent or average-based gains. If the corrected table reverses them, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an empirical superiority claim, so the comparison table must be trustworthy. The paper's own text in 'Comparison Results' asserts that on public datasets TSGaussian outperforms all baselines in PSNR, SSIM, and LPIPS and consistently achieves lower LPIPS, but Table 1 contains direct counterexamples. For garden, TSGaussian scores 27.93/0.942/0.049 while 3DGS scores 26.05/0.945/0.042, so 3DGS has higher SSIM and lower LPIPS. For bouquet, Gaussian Grouping has lower LPIPS (0.127) than TSGaussian (0.128). Several baseline entries are also implausible, e.g. SparseGS garden SSIM 0.502 at PSNR 22.16 and DNGaussian garden LPIPS 0.795 at PSNR 20.97, which suggests an evaluation-protocol bug rather than ordinary overfitting. Since the strongest claim 'outperforms state-of-the-art methods' rests entirely on this table, the evidence as presented does not establish it. The manuscript also reports no error bars, performs the ablation on a single scene, and does not make code or data verifiable, so the discrepancy cannot be resolved from the paper alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"TSGaussian proposes a target-specific 3D Gaussian Splatting framework for sparse-view novel view synthesis. The method uses YOLOv9 bounding boxes as prompts for SAM to obtain 2D target masks, a zero-shot tracker to maintain identity consistency, a compact 16-dimensional identity encoding per Gaussian with 2D and 3D grouping losses, a semantic-driven densification and pruning strategy, and multi-scale depth regularization borrowed from DNGaussian. The paper claims state-of-the-art performance on bear, bouquet, garden, and a newly collected citrus dataset, with the largest reported gain being a PSNR of 27.40 versus 18.90 for the best baseline on the citrus scene. The central contribution is empirical: the proposed combination of semantic constraints and depth priors is claimed to outperform existing methods for target-specific sparse-view reconstruction.","tokens_in":10109,"tokens_out":2882,"duration_ms":33298,"significance":"If the empirical claims were fully supported, this would be a useful contribution to sparse-view 3D reconstruction of specific objects, an application-relevant setting. The modular pipeline is sensible, the equations are coherent and mostly grounded in prior work (DNGaussian depth losses, Gaussian Grouping identity encoding), and the problem of target-specific rather than panoptic reconstruction is worth addressing. However, the significance is currently undermined by the evaluation: the main comparison table contains internal contradictions with the paper's own text, several baseline entries are implausible, the ablation is performed on a single scene, and no error bars or reproducible code verification are provided. The idea has promise, but the evidence as presented does not establish the headline claim.","major_comments":[{"comment":"The text in §5.2 states that on public datasets TSGaussian outperforms all baselines in PSNR, SSIM, and LPIPS and consistently achieves lower LPIPS, but Table 1 contradicts this claim. On garden, 3DGS has higher SSIM (0.945 vs 0.942) and lower LPIPS (0.042 vs 0.049) than TSGaussian; on bouquet, Gaussian Grouping has lower LPIPS (0.127 vs 0.128). Since the abstract and introduction rest on the claim of superior results over state-of-the-art methods, this table is load-bearing. Please correct the table or qualify the claims; also verify the entries, as several baseline numbers are implausible, e.g., SparseGS garden SSIM of 0.502 at PSNR 22.16 and DNGaussian garden LPIPS of 0.795 at PSNR 20.97, which suggest a systematic evaluation-protocol or metric-computation issue rather than ordinary overfitting.","section":"§5.2, Table 1"},{"comment":"The ablation study is reported on a single scene, with no error bars or repeated runs. The 'ALL' row (27.395 PSNR, 0.942 SSIM, 0.062 LPIPS) closely matches the TSGaussian row for the citrus dataset in Table 1 (27.40, 0.942, 0.063), suggesting the ablation is conducted on one scene only. The claimed conclusions about the importance of depth regularization (PSNR drop of 0.496) and semantic constraints (PSNR drop of 9.349) should be supported by ablations across all datasets, ideally with standard deviations, before the paper can claim consistent robustness.","section":"§5.3, Table 2"},{"comment":"The depth prior is a load-bearing component: the method relies on pretrained monocular depth estimates (Ranftl et al. 2020) after soft-hard and global-local normalization. The paper provides no validation of the depth estimator's accuracy on the evaluated scenes, despite the target objects being thin, translucent, or non-Lambertian (e.g., citrus seedlings, bouquet flowers). The ablation shows that removing depth regularization costs only 0.496 PSNR, so the benefit is modest; a failure analysis or at least a per-scene breakdown of the depth loss would be needed to establish that the depth prior is not introducing systematic biases.","section":"§3.4, Eqs. (5)-(11)"},{"comment":"The evaluation protocol for the baselines is under-specified in a way that affects the validity of Table 1. The text says 'all compared algorithms use the same semantic masks of specific objects,' but it is not stated whether evaluation metrics are computed on the full image, on a cropped target region, or with any masking. Baselines such as 3DGS and SparseGS do not use identity masks during training, so it is unclear what is being compared. Please specify exactly how PSNR/SSIM/LPIPS are computed, whether the same training views and iterations are used for all methods, and how the target-specific evaluation is performed for methods without semantic masks.","section":"§5.1, baseline protocol"}],"minor_comments":[{"comment":"The heading 'Ablation of Deep Regularization' should read 'Ablation of Depth Regularization' for consistency with Eq. (9) and the rest of the paper.","section":"§5.3, Table 2 heading"},{"comment":"The 3D regularization loss in Eq. (3) uses the notation F(e_j) log(F(e_j)/F(e'_i)) but does not define e'_i or clarify whether F outputs a probability distribution. If this is intended as a KL divergence, the normalization and summation order should be stated explicitly.","section":"Eq. (3)"},{"comment":"The section titled 'Limitations and Future Work' does not actually describe any limitations; it only outlines future directions. The paper should state what the method cannot handle, such as failure modes of the depth estimator or cases where YOLOv9/SAM mask predictions are incorrect.","section":"Limitations and Future Work"},{"comment":"The sparse-view sampling is described as 'one-third of the original views' split evenly into training and test sets, but no details are given on how the one-third views are selected, whether the split is random, or how many test views are used per scene. This makes the results difficult to reproduce.","section":"§5.1 experimental settings"},{"comment":"No error bars or standard deviations are reported for any quantitative result. Given the large variability commonly observed in sparse-view 3DGS training, a single run per scene is insufficient to support claims of consistent superiority.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is at an early stage and the main empirical claim is not yet supported by the evidence in Table 1. The discrepancies in the comparison table are serious enough that a simple rewrite of the text will not suffice; the authors need to rerun or re-report the experiments with a clearly specified protocol, correct the implausible baseline values, and provide multi-scene ablations. If the corrected results still show consistent gains, the paper may be suitable for publication after substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about this paper. The method is a sensible engineering integration: YOLOv9 boxes prompt SAM for masks, DEVA keeps IDs consistent across views, identity-encoded Gaussians (Gaussian Grouping style) plus DNGaussian-style depth losses supervise a target-focused sparse-view reconstruction, and pruning is driven by a semantic ROI. The agricultural application, citrus phenotyping, is a good fit. The second thing is that the main claim—outperforming SOTA on all metrics—is not supported by its own Table 1. On garden, 3DGS has higher SSIM and lower LPIPS than TSGaussian (0.945/0.042 vs 0.942/0.049). On bouquet, Gaussian Grouping has lower LPIPS (0.127 vs 0.128). The text says 'we outperform all baselines in terms of SSIM, LPIPS, and PSNR' and 'consistently achieves lower LPIPS'; the table directly contradicts that.\n\nThe math is coherent, mostly borrowed from established work, and I don't see a load-bearing theoretical flaw. The citations point to the right prior systems, and the semantic ablation is striking: removing semantic constraints costs 9.349 PSNR, which supports the importance of the masks. But the empirical validation has genuine soft spots. The table contains implausible baseline entries (SparseGS garden SSIM 0.502; DNGaussian garden LPIPS 0.795 at PSNR 20.97/SSIM 0.900), which suggests an evaluation-protocol bug rather than ordinary overfitting. No error bars are reported. The ablation runs on a single scene. The depth prior is never validated on these scenes, and removing depth costs only 0.496 PSNR, so that contribution is neither large nor well-calibrated. Pruning is underspecified—'floating mask' gets one sentence—and the code link isn't verifiable from the paper. The limitations section is boilerplate and doesn't acknowledge any of this.\n\nIf the numbers get fixed, the central story can hold: target-focused sparse-view reconstruction is useful and all components are publicly grounded. As submitted, the evidence does not establish the headline claim.\n\nMy take: this deserves a serious referee, but with an explicit instruction to redo the evaluation. The recipe is plausible, the problem is real, and the flaws are fixable. I wouldn't cite it until the table is corrected, but it's worth tracking if you work on sparse-view 3DGS or agricultural reconstruction.","headline":"Practical sparse-view target-specific 3DGS recipe, but Table 1 contradicts its own 'all metrics' claim, so the headline result needs a corrected evaluation before it can be believed.","tokens_in":10693,"tokens_out":3682,"would_cite":false,"duration_ms":36253,"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":"In TSGaussian, semantic masks and multi-scale depth priors guide 3D Gaussian Splatting to reconstruct a specified target from sparse 360° views with higher PSNR/SSIM and lower LPIPS than prior methods.","keywords":["3D Gaussian Splatting","sparse-view novel view synthesis","semantic segmentation","depth prior regularization","target-specific reconstruction","identity encoding","Gaussian pruning","monocular depth estimation"],"falsifier":"Run TSGaussian on a sparse-view scene dominated by thin or translucent elements, such as a wire sculpture or a glass bottle, and compare it against plain 3DGS with identical masks; if the depth-regularized version does not exceed 3DGS in PSNR, or its rendered geometry is visibly distorted, the claim that depth priors prevent geometric degradation is not supported.","tokens_in":9574,"feed_emoji":"🎯","tokens_out":11042,"duration_ms":95484,"temperature":0.7,"pith_summary":"TSGaussian claims that coupling semantic masks with monocular depth priors lets 3D Gaussian Splatting reconstruct a specified target object from sparse 360° views without geometric degradation. The paper's experiments report higher PSNR, SSIM, and lower LPIPS than four baselines on three public scenes and a new citrus dataset, with the largest gain an 8.50 dB PSNR improvement on the citrus scene. A sympathetic reader would care because sparse-view target reconstruction underpins agricultural phenotyping, augmented reality, and robotics, where full dense capture is costly. The core idea is to concentrate the Gaussian representation on the object of interest and use depth to keep geometry stable.","feed_headline":"Masks and depth priors lift sparse-view 3D reconstruction by 8.5 dB","feed_subtitle":"Combining semantic masks with monocular depth priors prevents overfitting in sparse-view novel view synthesis.","key_machinery":"The load-bearing machinery is the joint action of three modules. First, an identity encoding turns each Gaussian ellipsoid into a carrier of object identity, with a differentiable alpha-compositing renderer producing 2D semantic maps that are supervised by tracked masks and a K-nearest feature-distance loss. Second, a semantic-driven control and pruning step restricts densification to the 3D region of interest and removes Gaussians whose identity drifts, preventing background overfitting. Third, multi-scale depth regularization uses a monocular depth prior through a soft-hard depth loss that alternately fixes opacity and center, and a global-local depth loss that normalizes patches to correct small errors. Together these keep the Gaussian field focused on the target while preserving geometric plausibility in sparse-view settings.","core_discovery":"The central claim is that a target-specific 3D Gaussian field, driven by identity encodings learned from 2D masks and regularized by multi-scale depth losses, outperforms existing splatting methods when only one-third of the views are available. Concretely, TSGaussian assigns each Gaussian a compact identity code, renders semantics via alpha-compositing, and trains with a cross-entropy identity loss plus a K-nearest-neighbor consistency loss, so Gaussians cluster by object. It then densifies and prunes based on semantic attributes rather than purely positional gradients, focusing capacity on the target. A soft-hard depth loss and a global-local depth normalization loss, both anchored to a pretrained monocular depth estimator, keep the Gaussian centers and opacities geometrically consistent across unseen views. On the bear, bouquet, garden, and self-collected citrus scenes, the method reports PSNR/SSIM/LPIPS that beat 3DGS, Gaussian Grouping, DNGaussian, and SparseGS; for the citrus scene the reported PSNR is 27.40 versus 18.90 for the best baseline.","pith_inferences":["The 8.50 dB PSNR lead on the citrus scene may reflect the semantic pruning as much as the depth prior, since removing background Gaussians should reduce overfitting to sparse training views; a testable extension is to add the same pruning to DNGaussian or Gaussian Grouping.","Because removing depth regularization costs only 0.496 PSNR in the ablation, the depth prior may be acting as a mild stabilizer rather than the main driver of improvement; a fair reader could test whether a simpler smoothness regularizer yields a similar effect.","The pipeline depends on the detector and the tracker to propagate masks; if either fails on an unseen object class, the identity loss receives wrong supervision, so robustness to detection and tracking errors is an implicit boundary of the method."],"forward_implications":["With only one-third of the original views, the method reports PSNR/SSIM/LPIPS that beat 3DGS, Gaussian Grouping, DNGaussian, and SparseGS on the bear, bouquet, and garden scenes.","By pruning Gaussians outside the target's semantic region, the method reduces background allocation and redundant computations during optimization.","The identity encoding plus pruning lets the method output semantically clean masks of the target in novel views, not just render the object.","The citrus dataset demonstrates the method on slender plant structures where the paper reports a PSNR of 27.40 versus 18.90 for Gaussian Grouping."],"supporting_citations":[{"why":"Supplies the base 3D Gaussian Splatting representation and alpha-compositing renderer that the method extends.","marker":"(Kerbl et al. 2023)"},{"why":"Establishes the per-Gaussian identity encoding and semantic rendering idea that TSGaussian adapts to target-specific reconstruction.","marker":"(Ye et al. 2025)"},{"why":"Provides the soft-hard depth loss and global-local depth normalization scheme reused in the multi-scale depth regularizer.","marker":"(Li et al. 2024b)"},{"why":"Supplies the pretrained monocular depth estimator that generates the depth prior for each input view.","marker":"(Ranftl et al. 2020)"},{"why":"The Segment Anything Model that produces the 2D masks from box prompts for semantic supervision.","marker":"(Kirillov et al. 2023)"},{"why":"The YOLOv9 object detector that produces bounding box prompts for Segment Anything Model.","marker":"(Wang et al. 2024)"},{"why":"The tracking model that aligns identity masks across different views for consistent semantic supervision.","marker":"(Cheng et al. 2023)"},{"why":"The SparseGS sparse-view 360° Gaussian splatting baseline that motivates and benchmarks the sparse-view regularization.","marker":"(Xiong et al. 2023)"}],"fun_headline_variants":["TSGaussian: semantic and depth priors boost sparse-view 3D by 8.5 dB","Sparse views no problem: TSGaussian hits 8.5 dB gain with masks and depth","Target-specific Gaussian splatting: 8.5 dB sharper from sparse views","Depth and masks guide Gaussian splatting to 8.5 dB better sparse-view synthesis","Semantic masks plus depth priors: 8.5 dB sparse-view boost for targets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes the monocular depth estimates are accurate enough, after normalization, to serve as a reliable geometric prior for the target region; if they contain systematic errors on thin, transparent, or reflective structures, the depth losses will pull Gaussian positions and opacities to wrong places.","fun_headline_variants_meta":{"raw":{"variants":["TSGaussian: semantic and depth priors boost sparse-view 3D by 8.5 dB","Sparse views no problem: TSGaussian hits 8.5 dB gain with masks and depth","Target-specific Gaussian splatting: 8.5 dB sharper from sparse views","Depth and masks guide Gaussian splatting to 8.5 dB better sparse-view synthesis","Semantic masks plus depth priors: 8.5 dB sparse-view boost for targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000907,"raw_usage":{"total_tokens":3922,"prompt_tokens":992,"completion_tokens":2930,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":2811}},"tokens_in":608,"tokens_out":2930,"duration_ms":24005,"temperature":1.0,"reasoning_tokens":2811,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:24:52.452704+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TSGaussian on a sparse-view scene dominated by thin or translucent elements, such as a wire sculpture or a glass bottle, and compare it against plain 3DGS with identical masks; if the depth-regularized version does not exceed 3DGS in PSNR, or its rendered geometry is visibly distorted, the claim that depth priors prevent geometric degradation is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained monocular depth estimator that generates the depth prior for each input view."}],"review_version":1}