acceptodds
Under review as a conference paper at ICLR 2027

VisNet: Projective Geometry-View Fusion for Point Cloud Visibility

Abstract

Per-point visibility prediction from raw point clouds requires reasoning about both 3D geometry and view-dependent occlusion. Existing learning-based approaches primarily infer visibility from geometric representations conditioned on viewing direction, without explicitly modeling occlusion interactions between points that become locally coupled after projection into the image plane. We introduce projective geometry-view fusion, which combines 3D geometric context with explicit occlusion reasoning over projected point neighborhoods. VisNet implements this idea by constructing a view-dependent graph over projected neighborhoods and reasoning about possible occlusion through local message passing. The resulting view-specific representations are then fused with 3D geometric features through cross-cue interactions and neighborhood-level information exchange on the same projected graph. VisNet achieves state-of-the-art performance on sparse, noisy, and out-of-distribution point clouds, improving accuracy by 1.4-3.4% and mIoU by 3.0-6.9% over prior methods across 1k-32k inputs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Point Cloud Processing, Geometric Deep Learning, Visibility Prediction, Occlusion Reasoning, Computer Graphics, Projective geometry-view fusion

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_4GFhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Predict per-point visibilty from a raw cloud + calibrated camera. Main claim is 3D kNN misses the stuff that only shows up after you project (points far in 3D can overlap on the image). They build kNN on the view plane, point edges toward the camera by depth, GENConv on view feats, keep the same O-CNN geom branch as NeuVis, gate-fuse + Point Transformer on that graph, BCE. Train on ~38k cleaned ShapeNet, 24 views, raycast labels, 1k-32k pts. Compare HPR / HPRO / retrained NeuVis, also Blensor noise, ABC, some EinScan scans, runtime to 200k

Critical strengths
  • I buy the graph-on-the-viewplane idea, occlusion really is a 2d thing. and they reuse NeuVis’s O-CNN so Table 3 ablations are actually about the view graph/fusion, not a fancier 3D encoder
  • experiments are pretty complete for this problem. biggest jump vs NeuVis is in the sparse setting (+3.4 acc / +6.9 mIoU at 2k, Table 1). noise / ABC / directed vs undirected / runtime are all there. they even have a short limitations para which a lot of papers skip
Critical weaknesses
  • increment over NeuVis is not huge. its basically “stick a projected kNN GNN + gated fusion on the same geom encoder”. fine as engineering, but no analysis of when that kNN even contains the real occluder, no dumb z-buffer/splat baseline, and they always assume known K,R,t
  • Table 2 0% noise at 2k is 92.6, Table 1 is 96.7. I assume 0% still means Blensor? should just say that in the caption. also weird that HPRO loses to HPR everywhere, did they sweep the sphere radius or just default. downstream stuff is only qualitative. and GT is mesh raycast so you are partly just imitating a z-buffer from samples
Questions for authors
  1. Table 2 “0%” — clean ShapeNet or Blensor w/ no extra noise. why != Table 1
  2. HPR/HPRO hyps per density? swept inversion radius or defaults
  3. noisy extrinsics / unknown f, does the projected graph fall apart
  4. did you try a non-learned depth-buffer or spherical-flip+ztest on the same neighbors. i want to know how much is “use the view plane” vs the GNN
  5. EinScan — any quantitative number (agree with mesh visibility or whatever), figures only is a bit thin
3 comments

\textit{Due to the comment length limit, we have split our response into three parts. Please see our follow-up comment for the remaining responses. We thank you for your understanding.}

We thank the reviewers for their insightful and detailed comments. Here we address the main questions and concerns raised in the reviews:

Motivation and contribution: Our key observation is that spatial proximity in 3D does not necessarily reflect visibility relationships, as geometric proximity is not inherently correlated with occlusion. However, occluders and occluded points often become spatially adjacent when projected onto the image plane, revealing visibility interactions that are not apparent in 3D space. Based on this, we introduce VisNet, which makes these candidate interactions explicit through a graph over the original points, and combines the resulting view-dependent evidence with 3D geometric context. Our main contribution is a projection-induced formulation of visibility reasoning, where a depth-directed graph enables explicit information exchange between potential occluders and occluded points, including those far apart in 3D, and serves as a shared relational structure for both view-dependent occlusion reasoning and geometry-view fusion. By deliberately retaining the O-CNN geometry backbone (as in SOTA NeuVis), we show that the performance gains arise solely from our improved representation of visibility interactions and geometry-view fusion. We will sharpen this distinction in the revision and substantiate it with the neighborhood ablation below.

Projected-space vs. 3D kNN: Following this suggestion, we retrained a model using 3D kNN neighborhoods, achieving accuracies of 95.18%/96.05%/96.72%/96.86% at densities of 2k/8k/16k/32k, respectively, with mIoU values of 88.93%/91.22%/92.65%/93.24%. For reference, the projected-kNN model used in our VisNet achieves 96.7%/97.4%/97.8%/98.0% accuracy and 92.50%/93.88%/94.91%/95.48% mIoU (see Table 1), yielding gains of 1.52/1.35/1.08/1.14 percentage points in accuracy and 3.57/2.66/2.26/2.24 percentage points in mIoU, supporting the benefit of our projection-induced neighborhoods. We will include this comparison in the revision.

Camera calibration errors: We evaluated VisNet under individual errors in camera rotation, position, and focal length across 2k–32k points (recomputing visibility after each perturbation and testing nonzero rotation/position errors across five random seeds). Accuracy dropped by only 0.05–0.15 percentage points for a 5° rotation error, 0.20–0.40 percentage points for a position error of 5% of the object radius, and at most 0.032 percentage points when scaling horizontal and vertical focal lengths by up to ±10%. These results confirm VisNet's robustness to camera parameter errors within the tested ranges, and we will include the full results and settings in the revision. Quantitative diagnostics of EinScan scans: Since the underlying GT labels are unavailable for these real scans, we adopted an indirect approach to approximate the reference labels using an intermediate Screened Poisson reconstruction. Under this approximate evaluation protocol, VisNet achieved an average accuracy of 93.74%, an mIoU of 87.07%, and a visible F1-score of 91.97%, while NeuVis achieved only 83.11%, 70.95%, and 81.25%, respectively. We will include it in the revision.

Table 1 vs. Table 2: The apparent discrepancy in the reported values stems from differences in the underlying point-cloud distributions. Specifically, Table 1 reports results on uniformly sampled ShapeNet meshes, whereas Table 2 evaluates performance on simulated real scans, which retain the irregularities and non-uniformities characteristic of real-world point-cloud acquisition. And “0%” in Table 2 indicates that no additional noise is introduced beyond the artifacts inherent to the simulated scanning process. We will make this distinction clear in the caption and experimental description as suggested.

Depth-buffering and point-splatting baselines: As suggested, we evaluated a single-pixel point Z-buffer and fixed-radius disk splatting followed by a depth test, without neural networks or surface reconstruction. At 2k/8k/16k/32k point densities, the single-pixel baseline resulted in 42.33%/57.32%/65.90%/74.28% accuracy, while disk splatting improved this to 78.29%/81.23%/82.73%/84.13%, compared to VisNet, which achieves an accuracy of 96.7%/97.4%/97.8%/98.0% (see Table 1). The corresponding mIoU values are 26.37%/39.83%/48.20%/57.28% and 58.41%/62.94%/66.17%/68.61%, respectively (compared to 92.50%/93.88%/94.91%/95.48% for VisNet). Note that we empirically chose the disk radii that decrease from 13 to 7, 5, and 4 pixels as density increases. These results illustrate the importance of finite point support and show that the evaluated fixed-support projection rules incur substantial visibility errors. At 2k/8k/16k/32k points, the recorded runtimes are 0.42/0.78/1.76/3.49 ms per sample for single-pixel Z-buffering and 7.84/36.49/72.55/159.93 ms for disk splatting, compared with 27.40/32.01/39.44/57.15 ms for VisNet. Although single-pixel Z-buffering incurs relatively low computational overhead, it achieves substantially lower accuracy. Conversely, disk splatting is more computationally expensive and yields lower accuracy, making it a less favorable trade-off overall. We will add these accuracy and runtime results, together with the parameter-selection protocol and timing details, to the revision.

Mesh-raycast supervision vs. z-buffering: Although both approaches aim to achieve the same physical visibility objective, VisNet goes beyond fixed per-pixel depth comparisons by jointly reasoning about projected visibility interactions and the underlying 3D geometry. Furthermore, our previous experiments demonstrate that VisNet substantially outperforms conventional Z-buffering, highlighting the benefits of explicitly modeling these interactions.

Non-learned projected-neighborhood reasoning: We evaluated a fixed, non-learned Gaussian occlusion rule on exactly the same directed projected graph used by VisNet, with k=20 and identical edge orientation and deduplication. The heuristic aggregates evidence using projected distances and relative camera depths, achieving 79.23%/82.65%/83.59%/85.79% accuracy and 61.46%/65.87%/68.77%/71.79% mIoU at 2k/8k/16k/32k points compared to 96.7%/97.4%/97.8%/98.0% accuracy and 92.50%/93.88%/94.91%/95.48% mIoU for VisNet (Table 1). We also conducted an additional experiment by retaining a heuristic that uses up to 50 non-self neighbors among valid projections, which achieved 76.66%/80.08%/81.67%/82.83% accuracy and 55.74%/60.24%/62.54%/64.67% mIoU. This shows that explicitly using projected neighbors with this fixed rule is insufficient: even on the same graph, its visibility predictions remain substantially less accurate than VisNet’s.

Thin surfaces: We conducted an additional experiment using procedurally generated thin rectangular slabs with varying thickness and sampling density to explicitly evaluate performance on thin structures. We evaluated VisNet across slab thicknesses of 0.5%, 1%, and 2% (relative to the rectangle's side length), point densities (2k-32k), and noise levels (0.1%-1% of the BBox diagonal). Under noise-free conditions, accuracy and mIoU scaled with density and thickness, rising from 83.71%/71.67% (2k density, 0.5% thickness) to 97.72%/95.40% (32k density, 2% thickness). In an extreme stress test at 0.1% thickness, performance dropped to 60.18%-68.34% accuracy and 42.11%-51.57% mIoU across densities. Adding noise further degraded results: while 0.1% noise retained strong accuracy (up to 92.45% at 8k density, 2% thickness), higher noise levels caused sharp drops, falling to 62.93%-68.22% accuracy and 34.67%-43.62% mIoU at 1% noise. We will include these findings in the revision.

Known (K,R,t) requirements: VisNet is not restricted to a fixed camera configuration. It requires only a camera position and viewing direction to define a viewpoint. Although we adopt default (K,R,t) parameters in our experiments, these can be freely adjusted by users. The 2D projection plane and subsequent view-graph construction are then adapted accordingly, without requiring any changes to the underlying methodology.

Occluder coverage: As suggested, we evaluated the coverage (percentage of tested occluded points whose incoming graph neighbors include at least one sample on the surface blocking the point) of nearby occluding surfaces by graph neighbors. Across 2k-32k densities, occluder coverage increases from 80.23%-81.09% (k=10k=10) to 89.45%-90.39% (k=20k=20) and 91.68%-94.90% (k=30k=30-5050). However, higher kk yields negligible accuracy gains: at 32k density, k=20k=20 achieves 98.27% accuracy, while k=30k=30 and k=40k=40 add only 0.04 and 0.02 percentage points, and further increases degrade performance, while raising RTX 3090 inference time from 77.49ms (k=20k=20) to 91.47ms (k=30k=30), 109.16ms (k=40k=40), and 126.26ms (k=50k=50). Thus, k=20k=20 offers the optimal accuracy-efficiency trade-off (97.09%-98.27% accuracy). We will include this in the revision.

HPR/HPRO parameter selection: To ensure a fair comparison, we tuned HPR and HPRO across different point densities and used the best-performing configurations. We used the power-kernel parameter gamma values of -0.01/-0.005/-0.005/-0.001 for HPR and -0.005/-0.001/-0.0005/-0.0005 for HPRO at 2k/8k/16k/32k points, respectively; the HPRO runs use k=10 and a score threshold of 0.99. We will specify the search ranges, selection criterion, and exact configurations in the supplementary material. The ordering HPRO < HPR is also consistent with Table 1 of NeuVis.

Quantitative reconstruction: Across 360 objects (26 views each, 2k–32k densities), VisNet reduces mean squared Chamfer distance by 43.45%-54.77% and improves F-score by 2.22-2.39 percentage points over NeuVis.

We look forward to your feedback.

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

This paper presents a learning-based method for predicting point visibility from raw point clouds. The main idea is to explicitly model occlusion relationships in the projected view plane, which builds a view-dependent graph based on projected point positions and relative depth. The method is evaluated on ShapeNet at different point densities. VisNet improves over HPR, HPRO, and NeuVis in the reported experiments, with particularly noticeable gains on sparse point clouds.

Critical strengths
  1. The distinction between 3D geometric neighborhoods and projection-based neighborhoods makes sense for this task. Points that are far apart in 3D may still interact through occlusion when viewed from a particular camera. Building a graph in the projected space is a reasonable way to capture these relationships.
  2. The experimental results are encouraging. VisNet improves over NeuVis across all tested point densities on ShapeNet, with a particularly large improvement at 2k points. It retrains NeuVis on the same dataset and provide ablations for the geometry, view, and fusion components.
Critical weaknesses
  1. The results show that the proposed components are useful, but they do not directly establish how much of the improvement comes from the projected graph itself. A comparison that replaces projected kNN with 3D kNN while keeping the rest of the architecture unchanged would make the central claim more convincing.
  2. The paper compares against HPR, HPRO, NeuVis, and several reconstruction-based methods. However, since the proposed method relies heavily on projection-space reasoning, I would also like to see a comparison with direct depth-buffering or point-splatting approaches. These methods are discussed in the related work but are not included in the quantitative comparisons.
Questions for authors
  1. Could the authors provide a controlled comparison between projected-space kNN and 3D kNN using the same network architecture and approximately the same number of parameters? This would help clarify whether the improvements mainly come from the proposed neighborhood structure or from the additional network components.
  2. Have the authors considered evaluating against depth-buffering or point-splatting methods with appropriately selected point or splat sizes? How do these simpler projection-based approaches perform, particularly on sparse and noisy inputs, and how their runtime compares with VisNet.
2 comments

We thank the reviewers for their insightful and detailed comments. Here we address the main questions and concerns raised in the reviews:

Projected-space vs. 3D kNN: Following this suggestion, we retrained a model using 3D kNN neighborhoods, achieving accuracies of 95.18%/96.05%/96.72%/96.86% at densities of 2k/8k/16k/32k, respectively, with mIoU values of 88.93%/91.22%/92.65%/93.24%. For reference, the projected-kNN model used in our VisNet achieves 96.7%/97.4%/97.8%/98.0% accuracy and 92.50%/93.88%/94.91%/95.48% mIoU (see Table 1), yielding gains of 1.52/1.35/1.08/1.14 percentage points in accuracy and 3.57/2.66/2.26/2.24 percentage points in mIoU, supporting the benefit of our projection-induced neighborhoods. We will include this comparison in the revision.

Depth-buffering and point-splatting baselines: As suggested, we evaluated a single-pixel point Z-buffer and fixed-radius disk splatting followed by a depth test, without neural networks or surface reconstruction. At 2k/8k/16k/32k point densities, the single-pixel baseline resulted in 42.33%/57.32%/65.90%/74.28% accuracy, while disk splatting improved this to 78.29%/81.23%/82.73%/84.13%, compared to VisNet, which achieves an accuracy of 96.7%/97.4%/97.8%/98.0% (see Table 1). The corresponding mIoU values are 26.37%/39.83%/48.20%/57.28% and 58.41%/62.94%/66.17%/68.61%, respectively (compared to 92.50%/93.88%/94.91%/95.48% for VisNet). Note that we empirically chose the disk radii that decrease from 13 to 7, 5, and 4 pixels as density increases. These results illustrate the importance of finite point support and show that the evaluated fixed-support projection rules incur substantial visibility errors. At 2k/8k/16k/32k points, the recorded runtimes are 0.42/0.78/1.76/3.49 ms per sample for single-pixel Z-buffering and 7.84/36.49/72.55/159.93 ms for disk splatting, compared with 27.40/32.01/39.44/57.15 ms for VisNet. Although single-pixel Z-buffering incurs relatively low computational overhead, it achieves substantially lower accuracy. Conversely, disk splatting is more computationally expensive and yields lower accuracy, making it a less favorable trade-off overall. We will add these accuracy and runtime results, together with the parameter-selection protocol and timing details, to the revision.

We look forward to your feedback.

Thank you for the detailed rebuttal and additional experiments. The new comparison between projected-space kNN and 3D kNN provides useful evidence for the effectiveness of the proposed neighborhood structure. However, the response does not explicitly clarify whether the two variants use identical architectures. The disk-splatting radii are described as empirically selected, without sufficient details about the parameter-selection procedure. The accuracy-efficiency trade-off deserves a more careful discussion.

Reviewer_6Wt9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a method for predicting per-point visibility from an unstructured point cloud and a camera. Specifically, it combines camera-independent O-CNN geometry features with a view branch operating on a projected kNN graph. Edges are oriented from nearer to farther points to support potential occlusion reasoning. At last, gated feature fusion and a Point Transformer layer on the projected graph produce visibility predictions. The experiments validate the effectiveness of the proposed method.

Critical strengths
  1. The central geometric motivation is clear. Potential occluders can be distant in Euclidean 3D space but close in the projected view. Encoding these interactions explicitly is a sensible inductive bias for visibility prediction.
  2. The experiment shows that the method works for different numbers of points, which demonstrates its generalization.
  3. Several ablation studies are conducted to show the choice of directed vs. undirected, and the dual branches.
Critical weaknesses
  1. The whole pipeline is not novel to me compared to some existing projection-based baselines. Moreover, the comparison with such baselines is missing.
  2. Nearby projections can connect opposite sides of a thin surface, where small depth differences make near-to-far edge orientation sensitive to noise. Such mixing may blur visibility boundaries; controlled tests varying surface thickness, sampling density, and noise are needed.
  3. Evaluation is limited by object-level shapes. A few multi-object examples and higher point counts do not establish generalization to complex indoor or outdoor scenes with uneven sampling, missing regions, and long-range occlusions.
  4. It is not clear to me that how the proposed method handles behind-camera points, near-singular projections, and out-of-frustum points.
Questions for authors

Please refer to Weaknesses part.

2 comments

We thank the reviewers for their insightful and detailed comments. Here we address the main questions and concerns raised in the reviews:

Motivation and contribution: Our key observation is that spatial proximity in 3D does not necessarily reflect visibility relationships, as geometric proximity is not inherently correlated with occlusion. However, occluders and occluded points often become spatially adjacent when projected onto the image plane, revealing visibility interactions that are not apparent in 3D space. Based on this, we introduce VisNet, which makes these candidate interactions explicit through a graph over the original points, and combines the resulting view-dependent evidence with 3D geometric context. Our main contribution is a projection-induced formulation of visibility reasoning, where a depth-directed graph enables explicit information exchange between potential occluders and occluded points, including those far apart in 3D, and serves as a shared relational structure for both view-dependent occlusion reasoning and geometry-view fusion. By deliberately retaining the O-CNN geometry backbone (as in SOTA NeuVis), we show that the performance gains arise solely from our improved representation of visibility interactions and geometry-view fusion. We will sharpen this distinction in the revision and substantiate it with the neighborhood ablation below.

Projected-space vs. 3D kNN: Following this suggestion, we retrained a model using 3D kNN neighborhoods, achieving accuracies of 95.18%/96.05%/96.72%/96.86% at densities of 2k/8k/16k/32k, respectively, with mIoU values of 88.93%/91.22%/92.65%/93.24%. For reference, the projected-kNN model used in our VisNet achieves 96.7%/97.4%/97.8%/98.0% accuracy and 92.50%/93.88%/94.91%/95.48% mIoU (see Table 1), yielding gains of 1.52/1.35/1.08/1.14 percentage points in accuracy and 3.57/2.66/2.26/2.24 percentage points in mIoU, supporting the benefit of our projection-induced neighborhoods. We will include this comparison in the revision.

Thin surfaces: We conducted an additional experiment using procedurally generated thin rectangular slabs with varying thickness and sampling density to explicitly evaluate performance on thin structures. We evaluated VisNet across slab thicknesses of 0.5%, 1%, and 2% (relative to the rectangle's side length), point densities (2k-32k), and noise levels (0.1%-1% of the BBox diagonal). Under noise-free conditions, accuracy and mIoU scaled with density and thickness, rising from 83.71%/71.67% (2k density, 0.5% thickness) to 97.72%/95.40% (32k density, 2% thickness). In an extreme stress test at 0.1% thickness, performance dropped to 60.18%-68.34% accuracy and 42.11%-51.57% mIoU across densities. Adding noise further degraded results: while 0.1% noise retained strong accuracy (up to 92.45% at 8k density, 2% thickness), higher noise levels caused sharp drops, falling to 62.93%-68.22% accuracy and 34.67%-43.62% mIoU at 1% noise. We will include these findings in the revision.

Depth-buffering and point-splatting baselines: As suggested, we evaluated a single-pixel point Z-buffer and fixed-radius disk splatting followed by a depth test, without neural networks or surface reconstruction. At 2k/8k/16k/32k point densities, the single-pixel baseline resulted in 42.33%/57.32%/65.90%/74.28% accuracy, while disk splatting improved this to 78.29%/81.23%/82.73%/84.13%, compared to VisNet, which achieves an accuracy of 96.7%/97.4%/97.8%/98.0% (see Table 1). The corresponding mIoU values are 26.37%/39.83%/48.20%/57.28% and 58.41%/62.94%/66.17%/68.61%, respectively (compared to 92.50%/93.88%/94.91%/95.48% for VisNet). Note that we empirically chose the disk radii that decrease from 13 to 7, 5, and 4 pixels as density increases. These results illustrate the importance of finite point support and show that the evaluated fixed-support projection rules incur substantial visibility errors. At 2k/8k/16k/32k points, the recorded runtimes are 0.42/0.78/1.76/3.49 ms per sample for single-pixel Z-buffering and 7.84/36.49/72.55/159.93 ms for disk splatting, compared with 27.40/32.01/39.44/57.15 ms for VisNet. Although single-pixel Z-buffering incurs relatively low computational overhead, it achieves substantially lower accuracy. Conversely, disk splatting is more computationally expensive and yields lower accuracy, making it a less favorable trade-off overall. We will add these accuracy and runtime results, together with the parameter-selection protocol and timing details, to the revision.

Projection validity: All points lying outside the viewing frustum are considered invisible and are therefore excluded from subsequent processing.

Scene-scale generalization: Because VisNet relies solely on the projected plane and the scene's geometric features, it does not impose restrictions on indoor or outdoor environments. We will include additional visual results of objects from both indoor and outdoor environments.

We look forward to your feedback.

I thank the authors for the response. The rebuttal partially addressed my concerns, but I am still concerned about the comparison with learning-based projection methods. Moreover, additional indoor/outdoor object examples would not establish generalization to full scenes. Considering the limited methodological novelty, I am currently still maintaining my recommendation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.