acceptodds
Withdrawn from ICLR 2027

AnchorComplete: Recovering Complete Scene Meshes from Partial Point Clouds

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Most image-based 3D reconstructors prioritize reconstructing visible surfaces, producing partial point clouds with large holes and missing areas due to limited field of view and occlusions. To address this gap, we introduce AnchorComplete, a framework that recovers complete scene meshes from partial point reconstructions. Our design addresses two key challenges: anchoring generation to the observed geometry and scaling completion to diverse scene layouts. We introduce a partial-point conditioning module that guides both sparse-structure and sparse-latent generation using the input point cloud, together with Swin-DiT backbones that replace full self-attention with 3D windowed attention for efficient high-resolution generation. We instantiate these designs on top of the sparse latent representation of TRELLIS. The resulting model produces complete scene meshes that remain faithful to the observed geometry, providing a more complete digital representation of the physical environment. Experiments on Replica and SCRREAM show that AnchorComplete consistently outperforms strong baselines in complete scene reconstruction, particularly under sparse-view settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Scene Completion, 3D Generation

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

This paper focuses on recovering a complete scene mesh from a partial point cloud. Built on the TRELLIS architecture, the main technical contributions are: (1) partial-point conditioning — a frozen NOVA3R encoder produces global tokens injected via cross-attention, plus a local path that encodes the voxelized input into the SS latent, replaces the initial noise at visible cells, and concatenates the visibility mask at every step; (2) a Swin-DiT backbone (3D window attention implemented with masks), which makes a 32³ latent / 128³ occupancy grid fit in memory during training. The authors train on InternScenes and ScanNet++ and evaluate on Replica and SCRREAM. The results show a large lead over SG-NN, NKSR and Seen2Scene. The appendix adds a symmetric Chamfer, results for one checkpoint across seven input sources, and failure cases.

Critical strengths

1.Experimental results are much better than the baselines, and the design of how to inject partial scene point cloud, as well as the swin-dit are reasonable and align with the target.

  1. The ablation earns its place. Removing local injection is what breaks the model (CD_all 0.071 → 0.105), while dropping the global tokens costs much less (0.086), so per-voxel anchoring is clearly the load-bearing idea. Tab. 2b is candid about the Swin-DiT price, including the parts that don't flatter it: at 16³ full attention wins CD_all and [email protected], and Swin only pulls ahead at 2 cm.
Critical weaknesses

1.The comparison and discussion of closely related generative reconstruction methods are limited, particularly ReconViaGen and GenRecon. ReconViaGen also conditions TRELLIS generation on reconstruction information, while GenRecon extends an object-level generative prior to scene reconstruction. GenRecon is cited, but the paper offers neither a detailed methodological comparison nor an experimental one. These methods primarily use image conditioning rather than point-cloud input, and ReconViaGen targets objects, so they are not directly interchangeable baselines. I would like the authors to clarify the differences in conditioning, observation consistency and scene scaling, and add experiments where comparable inputs and implementations permit. This would make the contribution relative to this line of work clearer.

2.Table 2b supports the efficiency advantage of Swin-DiT, but the source of its quality gains could be better understood. Table 2a compares full attention at 16³ with Swin-DiT at 32³, changing both architecture and resolution. Reporting quality for Swin-DiT at 16³ would allow a same-resolution comparison with full attention. This would strengthen the ablation analysis without changing the existing evidence that Swin-DiT reduces computational cost and enables higher-resolution training.

Questions for authors
  1. What are the main differences from ReconViaGen and GenRecon in conditioning, observation consistency and scene scaling? Which follow from the input setting, and which are methodological?

  2. Could you add a scene-level comparison with GenRecon using the same input views and evaluation protocol, while specifying the conditioning information each method uses? If this is currently infeasible, what prevents it?

  3. How does Swin-DiT at 16³ compare in quality with full attention at the same resolution?

Reviewer_oqB9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AnchorComplete extends TRELLIS to complete scene meshes from partial point clouds through point conditioning and Swin-DiT backbones. The task is practically relevant, but the evidence for methodological novelty, advantages over closely related methods, and geometric consistency remains insufficient.

Critical strengths
  1. The method addresses a useful scene-completion problem, accepting partial point clouds from different reconstruction systems and producing meshes.
  2. Global point conditioning and local geometry injection form a reasonable design. The reported results demonstrate promising completion performance, while matched-resolution efficiency experiments support reduced training costs.
Critical weaknesses
  1. Insufficient coverage of closely related baselines. Many comparisons involve visible-surface reconstruction or image-conditioned generation rather than scene point-cloud completion. Although SG-NN and Seen2Scene are included, the absence of Points-to-3D and broader relevant completion baselines limits the conclusions.
  2. Unclear novelty and attribution of improvements. Local geometry injection closely resembles existing point-guided TRELLIS approaches, yet the differences are insufficiently discussed. Moreover, the quality ablation compares DiT and Swin-DiT at different resolutions, confounding architectural and resolution gains.
  3. Limited validation of geometric accuracy and generalization. The main metrics do not directly penalize erroneous geometry outside the input frustums. Although symmetric Chamfer distance is reported in the appendix, fine-detail recovery, generalization across scene scales, and placement accuracy in the original scene remain insufficiently evaluated.
Questions for authors
  1. Related methods and independent contributions: Why is Points-to-3D not directly compared, and how does the proposed local geometry injection substantively differ from its approach? Could the authors provide comparisons with relevant point-cloud completion methods under consistent training data and input conditions, together with matched-resolution ablations that isolate the contributions of conditioning and Swin-DiT?
  2. Resolution and generalization: Increasing latent resolution raises modeling dimensionality and may increase data requirements and sensitivity to noise. How does resolution affect cross-dataset generalization? Have the authors considered latent compression or multiscale representations to balance detail, efficiency, and generalization?
  3. Alignment and geometric bias: Voxelization, VAE encoding/decoding, and generation may introduce quantization errors and geometric shifts. How does the method preserve scale, position, and local surface alignment so that the output can be placed directly into the original scene?
Reviewer_TxwvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper builds on TRELLIS to generate complete indoor scene meshes from partial point clouds. It constrains generation through global point-cloud conditioning and local spatial injection, and adopts Swin-DiT to reduce the cost of high-resolution training.

Critical strengths
  1. The task has practical relevance, and the paper presents a complete system for scene mesh completion.
  2. Comparisons with Seen2Scene demonstrate advantages in completion quality and inference efficiency. The symmetric full-scene Chamfer results in the appendix further support the system’s effectiveness.
  3. The conditioning ablations and computational cost comparisons at matched resolutions provide evidence for the usefulness of the components and the training efficiency gains.
Critical weaknesses

1. The core conditioning mechanism closely resembles prior work, but the incremental contribution is not directly established.

The key operations in local spatial injection—including encoding the visible point cloud, using structure latents at observed locations and noise elsewhere, and concatenating a visibility mask—closely resemble the pipeline in Points-to-3D. Although the paper cites this work, it does not sufficiently explain the substantive differences or provide a direct comparison under comparable settings.

Moreover, local injection is the conditioning branch with the largest impact in Table 2a. The performance drop when removing it therefore primarily establishes the importance of this mechanism, rather than the paper’s contribution beyond prior work. The authors should clarify which designs are new and how much improvement these additions provide.

2. The backbone quality comparison confounds architecture and resolution.

Table 2b provides strong evidence for the training efficiency gains of Swin-DiT. However, Table 2a compares DiT at 16316^3 latent resolution with Swin-DiT at 32332^3. Improvements in fine-grained metrics therefore cannot be attributed independently to the backbone. DiT also achieves better results on CDall\mathrm{CD}_{\mathrm{all}} and F0.1F_{0.1}.

3. The primary metrics insufficiently penalize incorrect geometry in unobserved regions.

For both CDall\mathrm{CD}_{\mathrm{all}} and the F-scores, prediction accuracy or precision is evaluated only within the input-view frustums, while completeness or recall is evaluated over the full scene. Consequently, generating additional or incorrect surfaces outside the observed frustums may improve coverage without incurring a corresponding accuracy penalty. This limits the ability of the primary metrics to distinguish accurate completion from excessive geometry generation.

The symmetric full-scene Chamfer results in Appendix Table 4 still show an advantage over the baselines. However, the component ablations and sparse-view experiments primarily rely on the original metrics, leaving it unclear whether their conclusions hold when errors throughout the generated scene are penalized.

4. The sparse-observation robustness experiments lack generative completion baselines.

Figure 4 primarily compares against DA3 and NKSR, which reconstruct or fit visible surfaces. These methods lack a generative prior for recovering unobserved regions, so it is expected that the proposed method degrades more slowly on full-scene metrics as the number of input views decreases.

This experiment demonstrates the value of generative completion relative to visible-surface reconstruction, but does not sufficiently establish that the proposed designs provide stronger robustness to sparse observations. Including other generative reconstruction methods, or a conditioning baseline using the same generative prior, under the same view settings would help determine whether this advantage comes from the paper’s specific contributions.

5. The evaluation subsets and exclusion criteria limit conclusions about applicability.

The evaluation uses only 8 of the 18 Replica scenes and excludes SCRREAM samples whose videos focus only on a scene corner. The latter exclusion is particularly relevant because it directly concerns the limited coverage and incomplete observations that motivate the paper. However, sufficiently specific and reproducible selection criteria, along with an analysis of the exclusions’ impact, are missing.

Questions for authors
  1. What specific conditioning mechanisms are new relative to Points-to-3D, and which controlled comparisons establish the importance of these differences?
  2. After controlling for latent resolution, what independent effect does Swin-DiT have on completion quality?
  3. Do the key ablation and sparse-observation conclusions hold when using full-scene accuracy metrics and comparing against generative completion baselines?

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

Related papers

Discussion (0)

Sign in to comment.