acceptodds
Withdrawn from ICLR 2027

Joint Lossless Point Cloud Geometry and Attribute Compression via Causal Multiscale Coding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Geometry and attributes in colored point clouds are correlated, yet learned lossless codecs primarily exploit this dependency from geometry to attributes, leaving decoded attribute information unused for geometry probability estimation. A seemingly natural joint variant is to share geometry–attribute features throughout the encoder and decoder; however, this naive feature sharing does not necessarily improve overall lossless compression efficiency. We present a causal multiscale framework for joint lossless geometry and attribute compression. At each scale transition, decoded coarse-scale attributes condition finer-scale occupancy prediction; the finer-scale decoded geometry then provides support and context for attribute probability modeling. A shared joint processing unit (JPU) implements this interleaved cross-modal conditioning, with parameters reused across scales and modality-specific refinements for the two entropy models. On eight static point clouds from 8iVFB and Owlii, our method reduces geometry and total bitrate on every sequence relative to the separate Unicorn geometry and attribute codecs. On average, geometry, attribute, and total bitrates decrease by 5.0%, 0.7%, and 1.0%, respectively, while using 16.8% fewer trainable parameters. Under the same hardware and test settings, our method achieves 1.74 faster encoding and 1.69 faster decoding than the combined Unicorn codecs. Compared with the conventional codec G-PCC, the proposed method reduces the average total bitrate by 12.12%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point cloud compression, lossless compression, bidirectional geometry–attribute modeling, multiscale representation, conditional entropy modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RXZZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new method for joint point cloud geometry and attribute compression. The key idea is to leverage previously decoded coarse-scale attributes to predict finer-scale geometry during coarse-to-fine multiscale decoding, and then use the decoded geometry at the current scale to predict the corresponding attributes, thereby enabling bidirectional information flow between geometry and attributes.

Critical strengths
  1. The paper is clearly structured, and the proposed method is straightforward and easy to follow.

  2. The proposed method achieves better experimental results than the baselines.

Critical weaknesses
  1. The attribute compression in this paper primarily focuses on color information. However, point density is also an important characteristic of point clouds. Therefore, I believe the authors should evaluate the proposed method on point clouds with different density distributions, such as those in the KITTI [1] dataset, to further demonstrate its generalizability.

  2. The experiments are conducted on only eight point cloud samples, which I consider insufficient. The authors should conduct additional experiments on larger and more diverse datasets, particularly autonomous driving datasets such as KITTI.

  3. I believe the paper is missing two important baselines: V-PCC (https://github.com/MPEGGroup/mpeg-pcc-tmc2) and Draco (https://github.com/google/draco). Including these methods would provide a more comprehensive comparison with existing point cloud compression approaches.

  4. In the ablation study, the fully sequential Y → Co → Cg coding order achieves a lower bitrate than the default configuration. Why do the authors not adopt this setting in the final method?

  5. How does the proposed method compare with a combination of independently optimized geometry and attribute compression models, such as AnyPcc [2] for geometry compression and DALD-PCAC [3] for attribute compression? Would the proposed joint compression framework still achieve better performance than such a combination?

[1] Geiger A, Lenz P, Urtasun R. Are we ready for autonomous driving? the kitti vision benchmark suite[C]//2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012: 3354-3361.

[2] Wang K, Yi Q, Ye Y, et al. Anypcc: Compressing any point cloud with a single universal model[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 2972-2982.

[3] Fu C, Li G, Gao W, et al. DALD-PCAC: Density-Adaptive Learning Descriptor for Point Cloud Lossless Attribute Compression[J]. ACM Transactions on Multimedia Computing, Communications and Applications, 2026, 22(4): 1-23.

Questions for authors

Please see the "Critical Weaknesses" section

Reviewer_nmfiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a joint lossless point cloud geometry and attribute compression method based on causal multiscale modeling. It builds on Unicorn and adds a shared Joint Processing Unit (JPU) for interleaved cross-modal conditioning. The results show modest gains in both rate and speed.

Critical strengths
  1. The use of the multiscale hierarchy is reasonable. Cross-modal information is exploited without breaking the decoding order, and the motivation is clear.

  2. The paper reports a useful observation: "our ablation shows that such naive feature sharing does not improve overall lossless coding efficiency."

  3. Compared with Unicorn, the method is considerably faster and uses fewer parameters.

Critical weaknesses
  1. The bit rate improvement is small. Attribute bitrate improves by only 0.7% and total bitrate by 1.0%. With no significance analysis, it is hard to tell whether these gains are solid.

  2. The core innovation of this paper: "Decoded coarse-scale attributes provide additional information for fine-scale geometry entropy modeling." Direct experimental proof is lacking. Ablation experiments that only remove conditional inputs from coarse-scale attributes for fine-scale geometry prediction should be complemented.

  3. The experiments were conducted on only two human datasets (8iVFB and Owlii), and the comparison methods included solely G-PCC, Unicorn, and CNeT, missing comparisons with more recent studies, which makes the paper less convincing. Moreover, the reported bit rate of CNeT is significantly lower than that of this paper, and although CNeT has considerably higher decoding complexity, this discrepancy still warrants further investigation.

  4. The structures of the JE, Geometry refinement, attribute refinement, and ResBlock modules seem very straightforward. I did not find much novelty for the proposed architecture.

  5. Lidar point clouds (such as KITTI, scannet) should be tested to verify the generalization ability of the model.

  6. The figures lack visual polish. Fonts and figure layout should be improved.

Questions for authors
  1. In the ablation where coarse-scale attribute conditioning is removed, how much of the gain over Unicorn remains?

  2. Do you have results on more datasets and against more recent methods, separate or joint? Is the method still competitive?

  3. CNeT's total bitrate is clearly lower than yours. Why is that, and did you try to close the gap?

Reviewer_wihdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a learning-based lossless codec that jointly compresses geometry and attributes. The proposed method constructs a multiscale hierarchy using average pooling and performs coding in the order Al−1→Ol→AlA^{l-1} \rightarrow O^l \rightarrow A^l. Specifically, attributes at a coarser scale serve as context for occupancy coding at a finer scale, while the decoded occupancy is subsequently used for attribute coding. Although the idea is straightforward and easy to understand, I do not think the current experiments adequately demonstrate that the proposed coding order itself is responsible for the reported gains.

Critical strengths

None

Critical weaknesses

CW1. The proposed pipeline is evaluated using only one model, and the reported gains are marginal.

This paper does not introduce a new network architecture or probability model. The occupancy predictor, attribute predictor, parent-average update in Eq. (5), and zero-bit decoding of the last child are all inherited from Unicorn and Wang et al. (DCC 2023). Therefore, the main contribution lies in the proposed coding pipeline, specifically the Al−1→Ol→AlA^{l-1} \rightarrow O^l \rightarrow A^l coding order and feature sharing between the two modalities.

If the pipeline itself is the primary contribution, its effectiveness should be validated across more than one model. However, all experimental results are obtained using Unicorn as the backbone. Even with Unicorn, the gains are relatively small. The geometry bitrate reduction is only 0.028 bpp (from 0.550 to 0.522 bpp), corresponding to approximately 0.35% of the total bitrate of 7.925 bpp. Furthermore, for the Loot sequence, the proposed method actually produces a higher attribute bitrate than Unicorn.

CW2. It is difficult to allocate the geometry coding gains to attribute conditioning.

The occupancy network width is increased from 32 to 40 channels, and a 160-channel joint extractor is added. Consequently, the geometry coding path of the proposed model has greater network capacity than that of Unicorn.

The shared En/Dec variant uses the same causal coding order and the same Al−1A^{l-1} input, yet achieves a geometry bitrate of 0.460 bpp. This suggests that geometry coding performance may be primarily influenced by network capacity rather than attribute conditioning.

The isolated contribution of Al−1A^{l-1} is never evaluated. Based on the current results, it remains equally plausible that conditioning occupancy prediction on decoded attributes provides little or no benefit to geometry coding.

To substantiate the claimed contribution, I believe the following experiments are necessary: an ablation that removes only Al−1A^{l-1} from the joint extractor while maintaining the same network capacity, a capacity-matched separate (non-shared) baseline trained on the same data, and an evaluation of the proposed coding order using one or two additional backbones.

Questions for authors
  1. Could the authors provide the additional experiments discussed in CW2?

Specifically, please report an ablation that removes only Al−1A^{l-1} while maintaining the same network capacity, a capacity-matched separate baseline trained on the same dataset, and results obtained by applying the proposed coding order to at least one additional backbone.

These experiments would help determine whether the geometry coding gains arise from the proposed causal coding order or simply from the increased capacity of the geometry network.

  1. Could the authors provide a module-wise breakdown of the 44.73M parameters reported in Table 2?

The breakdown should include the joint extractor, the two refinement modules, the occupancy, luma, and chroma prediction networks, and the rounding-residual model, if used.

The paper attributes the reduction in model size to shared feature extraction and parameter sharing. However, Appendix A shows that the widths of the luma and chroma prediction networks are reduced from 128 channels to 40 and 80 channels, respectively.

Since the number of parameters scales approximately quadratically with channel width, I suspect that most of the parameter reduction comes from these narrower prediction networks rather than parameter sharing. In fact, the shared modules themselves introduce additional parameters. The same channel-width reductions could also be applied to separate codecs.

A module-wise parameter breakdown would help clarify where the reported parameter savings actually originate.

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

Related papers

Discussion (0)

Sign in to comment.