M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals
Vinícius da Silva, Isabelle Melo, Matheus Bessa, Guilherme Schardong, Luiz Schirmer, André Araújo, Nuno Gonçalves, Hélio Lopes, et al.
cs.CV · cs.GR · cs.LG
Quick verdict
Proposes nested-band residual SIRENs for noise-robust real-time SDFs; results support the claims, though fine mean CD trails iNGP and BACON is reimplemented.
Summary
The paper introduces M-plicits, a multiscale framework for neural signed distance functions (SDFs) that models a surface as a residual sum of SIREN MLPs trained under a nested-neighborhood scheme: a coarse network fits the full domain and each subsequent residual is supervised only within a narrow band around the previous zero-level set, with band width set adaptively from the maximum residual error at the input points. The authors argue the coarse level acts as a low-pass filter, giving robustness to noisy oriented point clouds, and add multiscale sphere tracing plus a GEMM-based analytical normal computation for real-time rendering. On Stanford and Thingi32, the coarse configuration achieves the best mean Chamfer distance (5.36E-05 vs iNGP's 6.87E-05) and the fine configuration the best median CD and IoU, using about an order of magnitude fewer parameters than iNGP; under 1% vertex noise, mean CD is 3.36E-05 versus 2.80E-04 for iNGP fine. Fine-config mean CD (6.57E-05) slightly trails iNGP fine (6.44E-05), which the authors disclose, and a Screened Poisson comparison shows SPSR wins clean IoU on a 10-shape subset. Main limitations: sensitivity to outlier contamination (IoU 0.002 on Lucy at 1% outliers due to the max-based band), a BANF comparison based on a weaker reimplementation, and no uncertainty reporting over repeated runs.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Modeling an SDF as a residual sum of SIRENs with nested-band supervision makes the coarse level act as a low-pass filter, yielding substantially better noise robustness than iNGP, BACON, and IDF (mean CD 3.36E-05 vs 2.80E-04 for iNGP fine under 1% vertex noise).
- On Stanford+Thingi32, M-plicits achieves the best mean Chamfer distance in the coarse configuration (5.36E-05) and the best median CD and IoU in the fine configuration, with roughly an order of magnitude fewer parameters than grid-based baselines.
- The nesting condition enables multiscale sphere tracing and adaptive marching cubes delivering real-time rendering (180 FPS coarse at 512²) and up to ~5× mesh-extraction speedup over a SIREN baseline.
- A GEMM-based analytical normal computation bypasses auto-differentiation and runs 2× faster than torch.autograd.
Strengths
Unusually honest results reporting
Sec. 4.1 explicitly discloses that fine-config mean CD (6.57E-05) trails iNGP fine (6.44E-05); S2.3 openly reports SPSR wins clean IoU on the 10-shape subset (0.910 vs 0.653); S2.4 states the BANF reimplementation trails BANF's published numbers by up to 2.9× at finer lattices and that BANF wins two clean shapes.
Informative two-axis ablation explaining why the design works
Suppl. S3.1 (Table S12) isolates residual composition and nested bands: without bands, mean Hausdorff is 3.6× worse under noise; without the residual sum, band-trained independent SDFs disintegrate outside the band (365–4,109 components, mean CD 2.8E-01 when extracted without culling), justifying both ingredients and the culling-based inference.
Strong efficiency and noise-robustness numbers
Concrete, measured systems contributions
Weaknesses
No statistical uncertainty reporting
BANF baseline is a reimplementation with a disclosed quality gap
Max-based adaptive band collapses under outliers
Speed claims are configuration-dependent
Evaluation protocol details incompletely specified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes M-plicits, a multiscale neural signed-distance-function (SDF) framework for reconstructing surfaces from oriented point clouds. Central thesis: represent the SDF as a residual sum of SIRENs (base + two residual MLPs, ω0 = 30/45/100) and supervise each residual only inside a narrow δ-band around the previous zero-level set (Eqs. 3–5). The coarse level is claimed to act as a low-pass filter so noise never becomes off-surface geometry; nesting enables multiscale sphere tracing, adaptive marching cubes, neural normal/texture mapping, and GEMM-based analytical normals. Headline claims: best mean Chamfer distance (CD) in the coarse configuration and best median CD and IoU in the fine configuration on Stanford + Thingi32; order-of-magnitude parameter reduction vs grid baselines; real-time rendering; up to 5× mesh-extraction speedup.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is M-plicits?
M-plicits is a neural implicit surface method that represents a signed distance function as a sum of residual SIREN networks trained in nested bands around progressively finer zero-level sets. It reconstructs surfaces from oriented point clouds and supports real-time rendering via a multiscale sphere-tracing algorithm, with the coarse network acting as a low-pass filter against input noise.
How does M-plicits handle noisy point clouds?
The coarse base network is frequency-limited (ω0=30), so it fits only smooth structure, and each residual network is supervised only within a narrow band around the previous (cleaner) surface — high-frequency noise never becomes off-surface geometry. Under 1% vertex noise the paper reports mean Chamfer distance 3.36E-05 versus 2.80E-04 for Instant-NGP fine. The method does remain vulnerable to spatial outlier points, which can inflate the adaptive band.
Is code available for M-plicits?
The paper states that code, models, and data will be released at github.com/dsilvavinicius/m-plicits. As with any 'will be released' statement, availability should be verified at the repository before relying on it.
What datasets and metrics were used to evaluate M-plicits?
Stanford scanning-repository models and Thingi32, plus a NeuS-based demo on DTU for multi-view images. Surfaces are extracted with marching cubes at 512³ resolution, 500K points are sampled, and quality is measured with L2 Chamfer distance and IoU (via PyTorch3D). Comparisons include Instant-NGP, BACON, IDF, and Screened Poisson reconstruction (supplementary).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (1-9, S10-S35) are images not analyzable from text extraction; qualitative claims (sharper iNGP details, blob artifacts, hole behavior in Fig. 4) are taken from captions only.
- The abstract is truncated mid-sentence in the extraction ('...best mean Chamfer distance'); remaining abstract content inferred from the body.
- The number of evaluation shapes for the main Table 2 is not stated in the extracted text; it may appear in a section not captured.
- Minor version-drift signal: S7.3 cites 'Fig. 13 in the main paper', absent from the extracted main text (Figs. 1-9).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.