acceptodds
Withdrawn from ICLR 2027

Decoupled Global-Local Implicit Representation for Continuous Nonlocal Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recovering signals from incomplete or noisy observations requires structural priors that remain effective across diverse sampling patterns. Nonlocal self-similarity (NSS) provides a powerful source of such prior knowledge by exploiting recurring local structures across spatially distant regions. Conventional NSS methods, however, typically rely on explicit neighborhood construction and grouping, making them less suitable for irregularly sampled observations and continuous coordinate queries. Implicit neural representations (INRs) offer a flexible alternative by modeling signals as continuous coordinate functions, but conventional INR formulations generally treat the signal as a single global implicit function and do not explicitly exploit recurring nonlocal structures. To bridge this gap, we propose a Decoupled Global-Local Implicit Representation (D-GLIR) for continuous nonlocal reconstruction. D-GLIR decomposes the implicit representation into shared local implicit bases and a coordinate-conditioned coefficient field. The local branch parameterizes reusable structural patterns over relative coordinates, while the global branch continuously maps spatial coordinates to the coefficients that combine these patterns at arbitrary query locations. Unlike conventional INRs that model the signal through a single global coordinate-to-value mapping, D-GLIR allows spatially distant regions to reuse the same local structural bases through different coefficient combinations. This introduces NSS as an intrinsic inductive bias of the representation while preserving its continuous coordinate-based nature, without explicit block matching, precomputed patch groups, or an additional NSS regularization term. Experiments on image inpainting, image denoising, off-grid function regression, point-cloud recovery, and weather data reconstruction demonstrate the effectiveness and flexibility of D-GLIR across diverse signal types, sampling patterns, and degradation conditions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Implicit neural representation, nonlocal self-similarity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_foMTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Decoupled Global-Local Implicit Representation (D-GLIR) for continuous non local reconstruction. D-GLIR represents a continuous local signal as g^(s,r)=∑k=1Kak(s)bk(r)\hat g(s,r)=\sum_{k=1}^{K}a_k(s)b_k(r). Where bk(r)b_k(r) is the k-th local basis function, and ak(s)a_k(s) is its location-dependent coefficient. Where local function parameterizes reusable structural patterns over relative coordinates, while the location dependent maps spatial coordinates to the coefficients that combine these patterns at arbitrary locations. Unlike conventional INRs that model the signal through a single global coordinate-to-value mapping, DGLIR allows spatially distant regions to reuse the same local structural bases through different coefficient combinations.

Critical strengths

1. Same formulation across different observation geometries

2. Consistent quantitative improvement across various experimental settings

Critical weaknesses

1. Insufficient evidence supporting the claimed NSS inductive bias

The central claim of this paper is "D-GLIR introduces NSS as an intrinsic inductive bias of the representation while preserving its continuous coordinate-based nature, without explicit block matching, precomputed patch groups, or an additional NSS regularization term". This claim is not sufficiently validated. In Eq. (4), the globally shared local bases resemble conventional methods or low-rank factorization, making it unclear whether the model specifically exploits nonlocal self-similarity or simply benefits from parameter sharing. The experiments lack controlled ablations, such as comparisons with independent or spatially partitioned dictionaries under matched model capacity, to isolate the claimed NSS mechanism. Consequently, the reported improvements cannot be confidently attributed to NSS, weakening the paper's central contribution

2. Insufficient justification of the Global-Local Decomposition

In Eq. (10), the final reconstruction is obtained as (f^(x)=∑k=1Kak(x)bk(0))(\hat f(x)=\sum_{k=1}^{K}a_k(x)b_k(0)). Since bk(0)b_k(0) is independent of the query location and the global coefficient field is evaluated through cubic interpolation, the final reconstruction is mathematically equivalent to interpolating an effective grid. Although the local branch may introduce useful structural constraints during training, the paper does not demonstrate whether these constraints provide a meaningful advantage over directly learning a continuous interpolated signal field.

Questions for authors

Q1: How does D-GLIR specifically exploits nonlocal self-similarity rather than benefiting from conventional methods?

Q2: Given Eq. (10), what is the advantage of the global-local decomposition over directly learning a continuous interpolated signal field?

Q3: How does the performance change with different numbers of shared local bases, patch sizes, and coefficient soft-thresholding values?

Q4: How does using globally shared local bases compare with using separate local bases for different spatial regions under similar model sizes?

Reviewer_6G34OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Authors present a model (D-GLIR) for reconstruction of signals from incomplete or noisy observations. They rely on the nonlocal self-similarity inductive bias (NSS), that structurally similar regions can aid in reconstruction of missing or noisy regions. They model the underlying signal with a local branch that maps relative coordinate to shared bases and global branch that maps absolute coordinated to coefficient that weight the bases. They utilize NSS "without explicit block matching, precomputed patch groups, or an additional NSS regularization". Bases and coefficients are fitted to the observed samples using reconstruction loss and regularizers (TV, coefficient sparsity, noise separation). The method is evaluated on regular grid: incomplete RGB and multispectral image completion, RGB image denoising, and irregular grid: synthetic and weather function regression/recovery and denoising, and point-cloud color regressing and denoising.

Critical strengths

The formulation is simple and unified across regular and irregular signals of various input and output dimensions.

In the experiments they consistently outperform the main (chosen) baseline CRNL, while being faster, having less parameters and a unified joint optimization, as opposed to fitting, partitioning, matching, and low-rank factorization reconstruction.

Critical weaknesses

Their representation is that of Dictionary Fields (DiF) [1] (not cited), and what D-GLIR adds is not isolated or tested. It has similar motivation: DiF's bases "model the commonalities of patterns that are shared across the entire signal domain", and factorized formulation: DiF models \hat{s}(x) = c(x)^T b(\gamma(x)) (Eq. 2): a coefficient field times shared bases. In Sec. 3, DiF states that "standard approaches with patch-wise basis … can be viewed as a special case of Eq. (2) when the coefficient field is piecewise-constant". D-GLIR with non-overlapping patches is exactly such a model.

NSS claim is insufficiently tested and not isolated. The paper claims the regularizers "are not responsible for the nonlocal behavior" (l.260–261), but the formulations and weights of the TV, sparsity and noise-separation terms are not given, nor ablated. The "inpainting" protocol appears easier than described, A.1 states "pixels are randomly retained" (l.661) but the Figs. 3, 4, 7, and 8 show individual channels being sampled and not full pixels missing. Moreover, structural (rectangular) holes missing would represent an inpainting task, where reusable local bases should matter most, and NSS effectiveness displayed.

The SIREN parametrization, mentioned in the main paper is only shown in the appendix (Table 10), and indicates no suitable \omega_0 frequency is used (depends on the signal and sparsity). SIREN-global improves significantly, while SIREN-local and -global perform worse than the baseline that with an appropriate frequency for bases, -local and -global could have better performance.

There is no resolution study and no visualization of the learned bases, no ablations, no qualitative figures in the main manuscript, no standard deviations despite small margins, multiple compared baselines seem not cited.

[1] Chen et al. Dictionary Fields: Learning a Neural Basis Decomposition. ACM TOG (SIGGRAPH) 2023.

Questions for authors
  1. How is D-GLIR different from Dictionary Fields [1]? What anchor spacing is used for gridded data, and how many anchors does each observation constrain? Can you show that overlapping local patches, rather than factorization, are what enable sparse recovery? eg compare with DiF, and D-GLIR with anchors spaced one patch apart (non-overlapping).

  2. Which sampling protocol is used (pixels or channels)? I suggest full (all channels) pixels/points excluded, structured mask inpainting, 1x1 patch experiments, ablations of each regularizer, visualizations of bases...

  3. What evidence shows non-local sharing?

  4. What are the exact regularizer forms, their weights, hyperparameters (K, patch size, SIREN \omega_0), and standard deviations over the five experiment trials?

Additional suggestions (not decision relevant): Arbitrary resolutions image queries experiments. Implicit shape function fitting (SDF/occupancy) to point clouds. Show qualitative figures with metrics and error maps into the main text and compact tables.

Reviewer_K5WFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes the Decoupled Global-Local Implicit Representation (D-GLIR), a continuous representation for reconstructing signals from incomplete or noisy observations. The authors argue that it exploits nonlocal self-similarity (NSS) without explicit patch matching, precomputed patch groups, or an NSS regularization term. D-GLIR represents the signal around each anchor location as a weighted combination of a small set of shared local basis functions. The local branch defines these bases over coordinates relative to the anchor, and the global branch produces location-dependent coefficients that determine how the bases are combined at each anchor. The reconstructed value at any location is read from the center of the local representation anchored there. Both branches are optimized jointly for each signal, using only its observations and no external training data.

In the main implementation, the local bases are learnable control lattices evaluated by bicubic or tricubic interpolation. The coefficients live on an anchor grid, are parameterized either directly or through an upsampled low-resolution latent with optional convolutional refinement, and are sparsified with learnable soft-thresholding. The method is evaluated on RGB and multispectral image inpainting, denoising of images, weather fields and point clouds, off-grid regression on synthetic functions and precipitation data, and point-cloud color recovery. It reports the best or near-best results on most benchmarks, against baselines that include tensor-factorization methods, SIREN, and the continuous NSS method CRNL.

Critical strengths

1. A simple, practical way to share local structure in continuous representations. The paper splits the representation into shared local basis functions and location-dependent coefficients, so similar local structures at distant locations can be represented with the same bases. Both components are optimized jointly for each signal from its observations alone, with no external training data. Because the formulation is coordinate-based, the same framework handles regularly gridded and irregularly sampled data.

2. Broad evaluation across signal types and tasks. The method is evaluated on RGB and multispectral inpainting; on denoising of images, weather fields and point clouds; on off-grid regression; and on point-cloud color recovery. Results are consistent across these settings, and D-GLIR is best or near-best on most benchmarks.

Critical weaknesses

1. Insufficient positioning against related prior work. The proposed factorization is closely related to patch dictionary learning, in which shared atoms represent local structures through patch-dependent sparse coefficients, as well as separable continuous-function representations. Although the paper cites earlier nonlocal sparse models and LIIF, it does not sufficiently distinguish its factorization from these established approaches.

2. The NSS claim is overstated. The only link between distant regions is a shared basis. Eqs. 26-28 hold for any shared basis, including a fixed one such as the DCT, so they do not show that the model exploits nonlocal self-similarity. No experiment shows that distant, similar regions actually receive similar coefficients.

3. The source of the gains is not isolated. In Table 10, the formulation with both branches as SIRENs reaches 19.82 dB, below plain SIREN at 24.30 dB. This suggests substantial sensitivity to the choice of function approximator, initialization, or optimization, despite the generality of the proposed factorization. D-GLIR also uses 8× more parameters than SIREN (2.114M vs 264K), and its coefficient grid scales with the number of data points.

Questions for authors

1. Because reconstruction uses only the center query, how does D-GLIR perform when trained only on center tuples, i.e., as a directly optimized interpolated coefficient grid? How does it perform with the regularization terms removed? These two ablations would isolate the contribution of the local neighborhood structure.

2. Why does SIREN-both underperform plain SIREN by roughly 4.5 dB, despite using the proposed global–local formulation? Can the authors clarify the source of this degradation?

3. Can you provide evidence that nonlocal reuse actually occurs, such as maps showing that distant, structurally similar regions receive similar coefficients?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.