acceptodds
Withdrawn from ICLR 2027

Point Clouds in the Frequency Domain: Color in Amplitude, Geometry in Phase

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Point clouds are widely used in vision applications, and colored point clouds, a more expressive 3D representation that encompasses both color and geometry, are becoming the norm in scanned and reconstructed data. Yet, a fundamental limitation persists: existing methods have largely been developed for point clouds with coordinates alone and simply append RGB values to each point as additional channels, leaving color and geometry entangled in the learned features. To address this, we pioneer the analysis of the amplitude and phase of colored point clouds under 3D Fourier decomposition. By reading color and geometry out of the amplitude and phase alone and by swapping the two components between point clouds, we find that color resides in the amplitude and geometry in the phase. Building on this finding, we propose an input scheme that takes amplitude and phase as inputs, so that color and geometry are learned independently. With this simple idea, our models perform favorably against recent methods on classification and segmentation of colored point clouds.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Colored Point Cloud, 3D Fourier Decomposition

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ary9OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a frequency-domain perspective on colored point clouds by voxelizing them and applying a 3D discrete Fourier transform to obtain amplitude and phase components for each of four channels (R, G, B, and an occupancy channel π). Through probing and swap experiments, the authors argue that color is encoded in the amplitude and geometry in the phase, and they introduce a Fourier-based input scheme that feeds amplitude and phase to separate encoders, whose fused features are used for classification and segmentation. Empirically, the method attains competitive or state-of-the-art results on DensePoint, ScanObjectNN, S3DIS, and Semantic3D, and ablations suggest the gains derive from the proposed representation rather than architecture tuning.

Critical strengths
  1. Clear motivation, pipeline figures, and problem framing; the Fourier decomposition and reconstruction are explained with sufficient mathematical and algorithmic detail for replication.

  2. The t-SNE visualization provides helpful intuition about why phase-derived features better capture shape semantics than raw XYZ pipelines.

Critical weaknesses

1.Heavy reliance on voxelization introduces quantization error and O(G^3) memory/compute; scaling to very large scenes or fine detail may be challenging.

2.Sensitivity to canonicalization: normalization removes translation/scale but not rotation; phase is known to be rotation-sensitive—rotation robustness is not analyzed.

  1. The “probabilistic” occupancy channel π is described as binary (0/1), which is somewhat inconsistent with its stated probabilistic interpretation; its design and thresholding are under-specified.

4.Reconstruction after amplitude/phase manipulation requires π-thresholding to suppress artifacts, suggesting residual misalignment that could limit applications beyond stylization/analysis.

5.The color-transfer metric compares only mean RGB distances, which ignores spatial color layout; stronger metrics (e.g., histogram distances, structural color similarity) would better test “color determined by amplitude.”

6.While the paper cites graph spectral methods and classic 3D FFT shape descriptors, more discussion contrasting voxel DFT (global, fixed basis) vs GFT (data-adaptive, no complex phase) would strengthen positioning, including robustness trade-offs.

Questions for authors

see weakness.

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

This paper investigates the frequency domain representation of colored point clouds, showing that Fourier amplitude primarily captures color information while phase preserves geometric structure. Based on this observation, the authors propose a Fourier-based input scheme with separate color and geometry encoders. The method is evaluated on point cloud classification and segmentation tasks, and the authors also show applications of style transfer.

Critical strengths

The frequency domain analysis provides an interesting perspective on separating color and geometry information in point clouds.

The method is evaluated across multiple classification and segmentation benchmarks, demonstrating competitive performance with recent baseline approaches.

Critical weaknesses

Insufficient experimental validation. The main claim regarding amplitude phase separation is primarily validated on synthetically colored objects. It is unclear how well this property generalizes to real-world point clouds with complex textures and noisy, or incomplete observations. More robustness experiments would strengthen the claims.

Limited performance improvements and qualitative evidence. The improvements over recent baselines are generally marginal (Table 4, 6, 8), raising questions about the practical benefits of the proposed representation. In addition, the style transfer results (Fig 8, 9) mainly demonstrate global color changes, with noticeable noise and limited preservation of meaningful local style patterns. The qualitative results are not sufficiently convincing.

Questions for authors

Can the authors validate the amplitude phase separation on real-world point clouds with more complex spatially varying colors and textures?

What is the additional computational cost of the Fourier-based input scheme compared with conventional approaches?

Can the authors provide additional experiments evaluating robustness to noisey and incomplete observations?

Reviewer_ohqSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the amplitude and phase of colored point clouds under a three-dimensional Fourier transform. It voxelizes point clouds into RGB and occupancy channels and argues that amplitude preferentially represents color, while phase preferentially represents geometry.

The authors investigate this claim through attribute-probing experiments and amplitude–phase swapping. They then introduce a dual-branch spectral encoder for classification and incorporate its global features into existing point-cloud segmentation networks.

Experiments cover object classification, part segmentation, and indoor and outdoor semantic segmentation. The paper additionally demonstrates point-cloud stylization by modifying amplitude while retaining the original phase.

Critical strengths
  • The paper addresses a clearly defined representation question and investigates it using complementary experiments. Probing evaluates attribute accessibility, whereas swapping examines how changing spectral components affects reconstructed attributes.

  • The downstream evaluation covers several datasets and tasks, including real scanned scenes. The improvement over the PTv3 backbone on S3DIS suggests that spectral features can provide useful complementary information beyond conventional point features.

  • The appendix provides useful mathematical qualifications of the proposed interpretation. In particular, it acknowledges that amplitude retains geometric autocorrelation information and that spatially varying color can affect phase, helping delineate the conditions under which the observed separation holds.

Critical weaknesses
  • The central claim is stronger than the supporting analysis warrants. The main text repeatedly describes color and geometry as being determined by amplitude and phase, respectively, whereas Appendix A establishes a more restricted relationship and explicitly allows geometric changes after amplitude swapping.

  • The principal probing and swapping experiments use objects painted with a single base color, with additional shading and noise. This setting closely follows the channel-scaling property already implied by Fourier linearity, while evaluating color transfer through mean RGB does not establish separation of spatially structured appearance from geometry.

  • The experiments do not sufficiently isolate the contribution of the proposed input representation. Table 5 does not clearly establish an architecture- and capacity-matched comparison across all input schemes, and the segmentation experiments introduce additional encoders and global features without corresponding voxel-domain or real–imaginary spectral controls.

Several improvements over strong baselines are small, including 0.05 percentage points in DensePoint classification accuracy over LitePT and 0.1 points in Semantic3D mIoU over PTv3. Without downstream uncertainty estimates and a runtime/memory comparison accounting for dense voxelization and the spectral encoders, the practical significance of these gains remains unclear.

Questions for authors
  1. How well does the proposed amplitude–phase separation generalize to complex scene-level point clouds containing multiple objects and spatially varying colors? Beyond downstream segmentation accuracy, can the authors provide scene-level probing or swapping experiments to validate this separation?

  2. For the style-transfer application, how does the proposed method compare with simple, learning-free color-transfer baselines, such as histogram matching or an affine color transformation applied directly to RGB values while keeping point positions fixed?

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

Related papers

Discussion (0)

Sign in to comment.