PAPER REPORTENAll readings ↗

PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Charles R. Qi; Hao Su; Kaichun Mo; Leonidas J. Guibas

Affiliations: Stanford University

Source: CVPR 2017 · ref-a6e65f34d4c161a2ba59 ↗ · Catalog record

Reading: 543 / 558 · 6 original figures & tables · ~19 min ·

1. Paper overview

In one sentence: Shared pointwise features and max pooling turn unordered 3D points into useful recognition representations, with learned alignment and a conditional critical-point account of robustness. e-taske-architecturee-theorye-classificatione-transformse-robustnesse-supplement

At a glanceWhat to know
Research problem
Source description

Voxel grids and rendered views can inflate point-cloud representations and introduce quantization artifacts. PointNet asks how a network can consume a set directly, ignore its arbitrary ordering, and retain geometric context for object recognition and per-point labeling. Permutation symmetry and robustness to geometric transformations are distinct requirements. e-taske-architecture

Core mechanism
Source description

A shared pointwise encoder plus symmetric max pooling supports classification; broadcasting the global descriptor back to point features extends the architecture to segmentation. e-architecture

A key reported resultModelNet40 object classification: 89.2 / 86.2

Overall / average-class accuracy (%). 9843 training / 2468 test CAD shapes; 40 classes; 1024 XYZ points per shape.

Handcrafted baseline: 77.4 / 72.6; Subvolume: 89.2 / 86.0; MVCNN: overall unreported / 90.1. PointNet matches Subvolume overall accuracy. MVCNN’s 90.1 is average-class accuracy and cannot be directly compared with 89.2. Input representations and view counts differ. e-classification

Reading caution
Reader analysis

The authors attribute the multiview classification gap to fine geometry captured by rendered images. The comparison changes representation and cannot isolate that explanation causally. e-classification

Core contributions

  • Source description

    A shared pointwise encoder plus symmetric max pooling supports classification; broadcasting the global descriptor back to point features extends the architecture to segmentation. e-architecture

  • Reader analysis

    Universal approximation assumes Hausdorff continuity and sufficient bottleneck width. Critical-set theory and controlled design comparisons help explain the architecture, without guaranteeing that a trained finite network solves every geometry task. e-theorye-pooling

Figure 2. One symmetric descriptor feeds object classification and per-point recognition. Original paper, p. 3 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Follow the blue path from left to right. The n-by-3 point array undergoes a learned coordinate transform, shared pointwise MLPs, a feature transform and another shared MLP before max pooling. Pooling removes the point axis and retains 1024 feature channels. The classifier maps that vector to k scores. For segmentation, follow the dashed branches: transformed 64-channel point features and the same global descriptor join into n-by-1088 features. Shared prediction layers output n-by-m scores. The two lower-left insets show T-nets predicting matrices that multiply coordinates or features; the arrows describe a forward computation rather than an iterative loop. e-architecturee-alignmente-parts

What it supports. The architecture separates ordering invariance from dense prediction. Max pooling gives a set-level summary, while retaining a branch for individual points lets each label depend on both that point’s features and the whole shape. The diagram shows pointwise processing and global aggregation, with no explicit neighborhood graph.

Where the evidence stops. This is the generic architecture. The evaluated part-segmentation variant has modifications deferred to the supplement. Learned alignment seeks transformation robustness; the picture alone does not prove exact rotation invariance.

2. Motivation

2.1 The problem and the proposed response

Source description

Voxel grids and rendered views can inflate point-cloud representations and introduce quantization artifacts. PointNet asks how a network can consume a set directly, ignore its arbitrary ordering, and retain geometric context for object recognition and per-point labeling. Permutation symmetry and robustness to geometric transformations are distinct requirements. e-taske-architecture

2.2 What this reading follows

A point cloud has geometry but no natural sequence. PointNet makes this mismatch an architectural constraint: every point passes through the same feature extractor, and a symmetric maximum summarizes the set. The resulting descriptor can classify an object or supply context to a predictor that labels each point. This reading follows that information flow, then separates three kinds of evidence: recognition accuracy, alignment ablations and corruption diagnostics. The critical-point visual explains why some changes leave the descriptor untouched, while also showing why that statement cannot promise protection against arbitrary missing geometry. The supplied CVPR paper is complete; its separately referenced supplement is outside this reading. e-taske-architecturee-theorye-classificatione-transformse-robustnesse-supplement

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryFoundational work
ArchitectureNot applicable
Prediction paradigmNot applicable
QuadrantNot applicable

This table preserves the labels recorded at reading time. The current major category is Components of WAMs. View the current classification.

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The foundational visual/3D representation placement is supported by the point-set encoder and recognition tasks. World-action architecture, prediction paradigm and quadrant are not applicable: there is no future/action predictor or controller. A unified recognition architecture does not establish a One Model world-action system. e-taske-architecture

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • Unordered XYZ points sampled from objects; optional additional channels.
  • Scene blocks use XYZ, RGB and normalized room coordinates.
  • Scores for k object classes.
  • An n-by-m array of per-point scores for m semantic or part labels.

4.2 Equations and their role

f(S)γ ⁣(MAXxiS{h(xi)})f(S)\approx\gamma\!\left(\operatorname{MAX}_{x_i\in S}\{h(x_i)\}\right)
In Eq. (1) and Theorem 1, S is the input set, h maps each point to K features, MAX takes componentwise maxima, and continuous gamma maps the pooled vector to a set-function value. Shared h and symmetric aggregation remove ordering dependence. e-architecturee-theory
Lreg=IAATF2L_{\mathrm{reg}}=\lVert I-AA^{\mathsf T}\rVert_F^2
Eq. (2) penalizes departures of the predicted feature-alignment matrix A from orthogonality, using identity I and the squared Frobenius norm. It is an added training penalty, not a hard constraint. e-alignment
CSTNS  f(T)=f(S),CSK\mathcal{C}_S\subseteq T\subseteq\mathcal{N}_S\ \Longrightarrow\ f(T)=f(S),\qquad |\mathcal{C}_S|\le K
Theorem 2 identifies critical set C_S and upper-bound set N_S for S; T is an intervening set and K the pooling width. Equality requires preserving critical responses and avoiding larger admitted responses; arbitrary deletions or outliers need not satisfy it. e-theorye-critical

5. Method in detail

5.1 Make ordering irrelevant before asking the network to recognize geometry

Source description

The source begins with an architectural requirement: a point set has no privileged order. Applying the same function h to every point makes point features move with their inputs under a permutation. Coordinatewise max pooling then removes that order entirely, because the maximum does not depend on which point was presented first. The prediction head sees only the pooled vector. Learned input and feature transforms address a different concern, alignment, and do not replace symmetric aggregation. Figure 5 tests several alternatives on ModelNet40: max pooling reaches 87.1% accuracy, compared with 83.8% for average pooling and 83.0% for attention sum. Those controls support this implementation choice; they do not imply that all other symmetric functions are incapable of representing sets. e-architecturee-alignmente-pooling

Figure 7. Sparse max-response witnesses explain a specific form of descriptor stability. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read each column vertically. The top row is the original point set S; the middle row shows critical points that supply the pooled responses; the bottom row depicts the upper-bound shape. Colors encode depth, not predicted parts or confidence. Theorem 2 says that sets T containing the critical set and contained within the upper-bound set preserve the output. Section 5.3 constructs the upper bound by admitting candidate points only when every component of their point feature is no larger than the corresponding pooled component. ‘Between’ therefore means set inclusion under a feature-response condition, not an interpolation of surfaces. e-criticale-theorye-architecture

What it supports. The sparse middle row illustrates how a descriptor can depend on only a subset of observed points. The dense bottom row illustrates a different freedom: some added points cannot change any maximum. This connects conditional robustness to the pooling operation and its bottleneck width, rather than to a learned geometric completion objective.

Where the evidence stops. The graphic is qualitative, and the theorem concerns the stated max-pooled point-function construction. These upper-bound shapes are not evaluated reconstructions. Arbitrarily deleting critical points or adding points with larger responses can violate the theorem’s conditions.

5.2 Give every point access to the same global context

Reader analysis

Classification can stop at a global descriptor, but segmentation must retain an association between features and input points. Figure 2 therefore keeps the transformed point features and broadcasts the pooled descriptor back to each one. The shared segmentation MLP receives both pieces, allowing a point label to depend on the whole object as well as the point’s own features. As a reader inference, this explains why the segmentation computation should follow a permutation of input points while preserving their corresponding labels. The source evaluates this design family on both ShapeNet parts and scanned-room semantic labels, with different inputs and protocols. Its part-segmentation implementation is explicitly modified from the generic figure, so the diagram alone is insufficient to reproduce Table 2. e-architecturee-partse-scenes

5.3 Treat critical-point robustness as a conditional statement

Reader analysis

For a fixed point function, each channel of the pooled vector has at least one point witnessing its maximum. Keeping suitable witnesses preserves that vector when other points are deleted. Adding points also preserves it when none exceeds an existing channel maximum. Theorem 2 expresses those conditions using the critical and upper-bound sets, and Figure 7 makes them visible. As a reader analysis, this is weaker than immunity to arbitrary corruption: a removed witness or a new high-response outlier can change the descriptor. Theorem 1’s approximation statement likewise assumes continuity and sufficient width rather than successful finite-data training. Figure 6 supplies empirical corruption evidence, but under particular sampling and training conditions. Together, theory and experiment explain useful stability without establishing universal robustness. e-theorye-criticale-robustness

5.4 Training and inference

During training

Source description

Supervised softmax training includes the feature-transform regularizer. Figure 2 specifies batch normalization with ReLU and dropout in the final classification MLP. The body omits optimizer, learning-rate schedule, epoch count, batch size, dropout rate and regularizer weight. e-alignmente-architecture

Source description

ModelNet40 training uses 1024 area-weighted mesh-surface samples, unit-sphere normalization, random up-axis rotations and zero-mean Gaussian coordinate jitter with standard deviation 0.02. Scene training samples 4096 points per block. e-classificatione-scenes

During inference

Source description

A forward pass predicts alignment matrices, point features, pooled context and labels. Scene evaluation processes all points under the cited k-fold protocol. A separate detection extension uses connected components of predicted semantic regions for proposals. No dynamics rollout or controller is defined. e-architecturee-scenese-detectione-task

5.5 Implementation flow

  1. Align and encode each point

    An input T-net predicts a 3-by-3 coordinate transform. Shared MLP(64,64) features pass through a 64-by-64 feature transform and shared MLP(64,128,1024). Each shared MLP processes points with identical weights. e-architecturee-alignment

  2. Pool across the unordered set

    Coordinatewise max pooling reduces n-by-1024 features to a 1024-dimensional global descriptor. Symmetry makes the descriptor invariant to point order. MLP(512,256,k) produces object-class scores. e-architecture

  3. Restore context to point predictions

    The segmentation branch concatenates each transformed 64-dimensional point feature with the global descriptor, forming 1088 channels per point. Shared prediction layers produce semantic scores. This broadcast is feed-forward, without temporal feedback or action execution. e-architecturee-task

6. Experiments & results

PointNet learns directly from unordered 3D points: shared pointwise networks extract features, coordinatewise max pooling summarizes a shape, and classification or segmentation heads predict labels. Learned alignment improves accuracy, while a critical-point analysis explains conditional robustness. The evidence establishes an efficient geometric recognition representation, with qualifications about input modalities, missing supplements and corruption settings.

6.1 Read the original evidence

Table 1. PointNet matches the listed Subvolume overall accuracy using point input. Original paper, p. 5 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the two accuracy columns separately. PointNet’s 86.2 is average-class accuracy, whereas 89.2 is overall accuracy. The Subvolume row reports 86.0 and 89.2 in those same columns. MVCNN reports 90.1 only under average-class accuracy; the dash in its overall column must remain missing. The input and view columns show that these are different representations and aggregation settings, not an otherwise identical encoder swap. The surrounding experiment text supplies the PointNet protocol: ModelNet40’s 9843 training and 2468 test shapes, with 1024 mesh-surface samples per object normalized to a unit sphere. e-classification

What it supports. The point-based network reaches the same 89.2% overall accuracy as Subvolume in this table and exceeds the authors’ handcrafted baseline at 77.4%. It remains below MVCNN on the comparable average-class metric, 86.2% versus 90.1%. The table supports competitive recognition without establishing the best result across every representation.

Where the evidence stops. No uncertainty is reported here. Different input modalities and view counts prevent a controlled attribution of the differences to architecture alone. MVCNN’s unreported overall accuracy cannot be reconstructed from its class-average entry.

Table 2. Mean part-label overlap improves, with several category-level exceptions. Original paper, p. 6 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start with the mean column, then scan categories to see whether that aggregate improvement is uniform. The entries are point-based IoU percentages, not object-classification accuracies. The source computes each part’s intersection over union, assigns one when both predicted and ground-truth part sets are empty, and averages part IoUs within a shape. Per-category values average those shape scores. The top row gives category shape counts, exposing substantial dataset imbalance. Keep the reported mean column as supplied: the body does not explicitly describe its weighting, so do not silently replace it with an unweighted average of the category columns. e-parts

What it supports. PointNet reports 83.7% mean IoU against 81.4% for Yi and 79.4% for the authors’ 3DCNN. The calculated gain over Yi is 2.3 percentage points. It does not win every column: car is 74.9 versus 75.7, and motorbike is 65.2 versus 70.6.

Where the evidence stops. The supplied body omits exact part-dataset split counts and defers the modified segmentation architecture to the supplement. This table evaluates the part benchmark; it is not the separate simulated partial-scan robustness result.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
ModelNet40 object classification

9843 training / 2468 test CAD shapes; 40 classes; 1024 XYZ points per shape.

89.2 / 86.2

Overall / average-class accuracy (%)

Handcrafted baseline: 77.4 / 72.6; Subvolume: 89.2 / 86.0; MVCNN: overall unreported / 90.1.

PointNet matches Subvolume overall accuracy. MVCNN’s 90.1 is average-class accuracy and cannot be directly compared with 89.2. Input representations and view counts differ. e-classification

ShapeNet part segmentation

16881-shape dataset, 16 categories, 50 parts; point-level evaluation. Exact train/test counts are unstated in the body.

83.7

Reported mean IoU (%)

Yi: 81.4; authors’ 3DCNN: 79.4.

The calculated gain over Yi is 2.3 percentage points. Part IoUs are averaged within shapes; empty-union parts receive one. Mean-column weighting is not explicitly described in the body. Category performance is not uniformly better. e-parts

Stanford scene semantic segmentation

Six areas / 271 rooms / 13 labels; 1 m by 1 m blocks; cited k-fold protocol.

47.71 / 78.62

Mean class IoU / overall point accuracy (%)

Handcrafted-feature MLP: 20.12 / 53.19.

These are scene-labeling results. The body does not explicitly name k or list folds. e-scenes

Scene 3D object detection

Segmentation-derived connected-component proposals; table, chair, sofa and board categories.

24.24

Mean AP at 3D volume IoU 0.5 (%)

Armeni et al.: 18.22.

Mean AP improves, but sofa AP falls to 4.76 from 6.78. Supplementary proposal details are needed for reproduction. e-detection

ModelNet40 pooling comparison

Section 5.2 design-analysis models in Figure 5.

Max pooling: 87.1

Classification accuracy (%)

Average pooling: 83.8; attention sum: 83.0; LSTM: 78.5; sorted MLP: 45.0; unsorted MLP: 24.2.

Supports max pooling among these tested implementations, without establishing superiority over every symmetric architecture. e-pooling

ModelNet40 alignment ablation

Table 5 transform variants on the test set.

Both transforms with regularization: 89.2

Overall classification accuracy (%)

None: 87.1; input only: 87.9; feature only: 86.9; regularized feature only: 87.4.

Feature transformation alone hurts; regularization helps. The combined setting gains 2.1 percentage points over no transforms, without reported uncertainty. e-transforms

Classification model complexity

Table 6 architectures; Subvolume and MVCNN aggregate multiple rotations or views.

PointNet: 3.5M / 440M

Parameters / FLOPs per sample

Vanilla: 0.8M / 148M; Subvolume: 16.6M / 3633M; MVCNN: 60.0M / 62057M.

Operation counts favor PointNet under these configurations; they are not standardized hardware latency measurements. e-efficiency

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Table 5. Alignment helps most when the coordinate and regularized feature transforms are combined. Original paper, p. 7 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Use the no-transform row as the reference, then distinguish coordinate alignment from feature alignment. The input matrix is 3-by-3; the feature matrix is 64-by-64. Feature transformation without the regularizer drops below the reference, while the next row adds the orthogonality penalty from Eq. (2). The final row says only ‘both’; the adjacent paragraph specifies that the best configuration combines both transformations with regularization. Read that prose together with the table rather than interpreting the short row label as a complete training specification. All entries are overall classification accuracy on ModelNet40. e-transformse-alignmente-architecture

What it supports. Input alignment raises accuracy from 87.1% to 87.9%. The unregularized feature transform reaches 86.9%, rising to 87.4% with regularization. Both transforms with regularization reach 89.2%, a calculated 2.1-point gain over the no-transform network. This supports a benefit from the combined design under the reported experiment.

Where the evidence stops. The table lacks uncertainty and a complete factorial comparison. It does not establish statistical significance, uniquely attribute the combined gain, or guarantee geometric invariance. The regularization coefficient is absent from the supplied main text.

Figure 6. Corruption tolerance has limits that depend on how points are sampled and altered. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read these as three separate interventions. On the left, the horizontal axis is the fraction of missing points; the blue and red curves distinguish furthest and random sampling. In the middle, uniformly scattered unit-sphere outliers vary in proportion, and the legend distinguishes XYZ input from XYZ plus density. On the right, independent Gaussian coordinate perturbations grow in standard deviation. The vertical axes all measure overall ModelNet40 accuracy, but their displayed ranges differ. The accompanying text identifies the Figure 5 max-pooling architecture for these tests; do not assume that every curve starts from the full transformed model’s Table 1 score. e-robustnesse-pooling

What it supports. The curves support tolerance to moderate corruption alongside pronounced failure under stronger corruption. Furthest sampling retains accuracy better than random sampling at severe deletion. Density does not improve every outlier setting. Together, the panels qualify the broad robustness claim by showing dependence on sampling, input features and corruption magnitude.

Where the evidence stops. The outlier experiment assumes exposure to outliers during training. These plots do not establish robustness to arbitrary sensor failures or adversarial point removal, and the source does not provide error bars or a complete corruption-training recipe.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

The authors attribute the multiview classification gap to fine geometry captured by rendered images. The comparison changes representation and cannot isolate that explanation causally. e-classification

Reader analysis

Robustness depends on corruption and sampling. Outlier tests expose training to outliers; the theorem protects only intervening sets containing critical points. Figure 6 shows substantial degradation under severe deletion, outliers or Gaussian noise. e-robustnesse-theory

Reader analysis

Result and ablation tables provide point estimates without run-to-run uncertainty. Modified segmentation, T-net and detection details are referred to separate supplementary material, limiting implementation-level conclusions. e-classificatione-partse-scenese-transformse-supplement

7.2 Questions for discussion

  1. How much geometry needed for dense labeling can survive a fixed-width max-pooled descriptor?
  2. How does robustness change when corruptions remove critical points rather than random points?

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reconstruction needs the specified sampling, normalization, augmentations and task-specific label conventions, plus missing training hyperparameters and supplementary architectures. Simulated Kinect partial scans and real scanned rooms are distinct evaluations. e-classificatione-partse-scenese-supplement

Reader analysis

For runtime replication, the source names TensorFlow and a “1080X GPU” exactly, without software version or sufficient timing configuration. That hardware label should not be silently reinterpreted. e-efficiency

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Proposed check: isolate regularization from alignment capacity

Train the five Table 5 configurations with the same ModelNet40 split, cached point samples, augmentation distribution, classifier widths and optimization budget. Predeclare the otherwise missing hyperparameters and run five matched seeds. Record overall accuracy and the feature matrix’s orthogonality penalty for each run. Compare the feature-only model with and without regularization before interpreting the combined-transform model. A reproducible gain accompanied by smaller orthogonality deviation would support the proposed stabilizing role; no gain across matched runs would weaken the single-table conclusion. Report all runs and distinguish the regularization contrast from the extra capacity introduced by adding transforms. This is a proposed experiment, not a reproduced result. e-classificatione-alignmente-transforms

Check 2: Proposed check: distinguish critical-point stability from random robustness

Use the vanilla max-pooling network with fixed weights and evaluation settings. For each cloud, retain one maximizing point per feature channel, taking the union across channels with deterministic tie handling. Delete only other points and test whether the pooled descriptor remains unchanged within a predeclared floating-point tolerance. Compare equal-size random deletions and deletions targeting the retained witnesses. Then repeat with the full network, first holding its original predicted transforms fixed and then recomputing them, to separate point-function stability from input-dependent alignment. A descriptor change after deleting only non-witness points in the fixed-function condition would falsify the claimed implementation invariant. This is a proposed diagnostic; no run is claimed. e-theorye-criticale-alignmente-architecturee-efficiencye-robustness

8.3 Reading coverage

Visual audit: All nine supplied PDF pages were rendered and visually inspected, as were all six final original crops. Page 1 verifies title, authors, affiliation and the CVF edition watermark; pages 2–5 cover task definitions, architecture, alignment equations and theorem conditions. Pages 5–8 supply every retained training, evaluation, numerical, robustness and compute detail, including the Figure 5 accuracy table absent from the text extraction. Page 9 contains references. Figure 2 branch directions were checked against Section 4.2 and Eqs. (1)–(2); no claim-relevant conflict was found. Table 5’s abbreviated ‘both’ row was checked against its adjacent regularization explanation. Figure 7 was read with Theorem 2’s set-inclusion condition. Figures 1–7 and Tables 1–6 were inspected; six are included as crops. No separate supplement, code or additional edition was inspected.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9. Appendix coverage: not present.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Abstract
  • 1. Introduction
  • 2. Related Work
  • 3. Problem Statement
  • 4. Deep Learning on Point Sets
  • 4.1. Properties of Point Sets in Rn
  • 4.2. PointNet Architecture
  • 4.3. Theoretical Analysis
  • 5. Experiment
  • 5.1. Applications
  • 5.2. Architecture Design Analysis
  • 5.3. Visualizing PointNet
  • 5.4. Time and Space Complexity Analysis
  • 6. Conclusion
  • Acknowledgement
  • References

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Separate supplemental material availability has not been fully verified.
  • All four supplied text chunks and all nine PDF pages were read. Direct visual inspection addressed the extraction’s figure omission.
  • No separate supplement or appendix was supplied. Supplementary proofs, detailed architectures and additional application results were not read. Code was not inspected and experiments were not reproduced.
  • Edition/identity note: the inspected artifact is the CVF CVPR Open Access PDF. Its watermark states identity with IEEE Xplore except for the watermark; no numbered revision or revision date is displayed. Title and authors match the catalog. The title page abbreviates the first author as Charles R. Qi. The catalog BibTeX gives pages 77–85, while this PDF prints 652–660. The discrepancy is preserved without changing the catalog or inferring another edition.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityPDF p. 1, title block, equal-contribution footnote and CVF watermark; pp. 1–9, printed folios 652–660Inspect

The title and four authors match the catalog author string. Stanford University is the affiliation; Qi and Su have equal-contribution marks. The watermark identifies the CVPR Open Access version and states that it differs from the IEEE Xplore version only by that watermark. Printed pagination differs from the supplied BibTeX pagination.

Go to primary source ↓
e-taskPDF pp. 1–2, Introduction and Section 3; p. 3, Figure 2Inspect

The network consumes unordered point sets and produces object-class scores or per-point semantic scores. Its applications are geometric recognition, without future-state or action outputs.

Go to primary source ↓
e-architecturePDF p. 3, Figure 2 and caption; p. 4, Section 4.2, Eq. (1) and Local and Global Information AggregationInspect

Shared pointwise MLPs, coordinate and feature transforms, and max pooling form the classification path. The segmentation branch concatenates 64-dimensional point features with the 1024-dimensional global feature. The caption specifies batch normalization with ReLU and dropout in the final classification MLP.

Go to primary source ↓
e-alignmentPDF p. 4, Section 4.2, Joint Alignment Network and Eq. (2)Inspect

T-nets predict matrices applied directly to coordinates and features. A squared Frobenius penalty on I minus A times its transpose is added to the softmax training loss to encourage the feature transform toward orthogonality.

Go to primary source ↓
e-theoryPDF pp. 4–5, Section 4.3, Theorems 1–2 and their explanationsInspect

The approximation theorem concerns Hausdorff-continuous functions on fixed-cardinality sets in a bounded domain, given enough pooling channels. The critical-set theorem states unchanged output for sets between the critical and upper-bound sets, with at most K critical points. Theorem 1's proof is referred to supplementary material.

Go to primary source ↓
e-classificationPDF p. 5, Table 1 and Section 5.1; p. 6, continuation of 3D Object ClassificationInspect

ModelNet40 has 9843 training and 2468 test shapes. Inputs are 1024 area-weighted mesh-surface samples normalized to a unit sphere, augmented with up-axis rotations and Gaussian jitter of standard deviation 0.02. Table 1 reports PointNet 86.2% average-class and 89.2% overall accuracy, Subvolume 86.0/89.2, the handcrafted baseline 72.6/77.4, and MVCNN 90.1 average-class accuracy with no overall entry.

Go to primary source ↓
e-partsPDF p. 6, Table 2 and Section 5.1, 3D Object Part Segmentation; p. 5, Figure 3Inspect

ShapeNet part data comprise 16881 shapes, 16 categories and 50 parts. Table 2 gives mean point IoU of 83.7 for PointNet, 81.4 for Yi and 79.4 for the authors’ 3DCNN. Empty-union part IoU is counted as one; part IoUs are averaged within shapes. PointNet car/motorbike IoUs are 74.9/65.2 versus Yi's 75.7/70.6. Modified segmentation details are deferred to the supplement. Partial examples use simulated Kinect scans from six viewpoints; the text reports a 5.3% mean-IoU loss without a separate numerical table.

Go to primary source ↓
e-scenesPDF p. 6, Table 3 and Semantic Segmentation in Scenes; p. 7, continuation and Figure 4Inspect

The Stanford dataset has six areas, 271 rooms and 13 classes. One-square-meter blocks use XYZ, RGB and normalized room location, with 4096 points sampled per training block and all points tested. A k-fold protocol is cited without an explicit k in this body. Table 3 gives 47.71 mean IoU and 78.62 overall accuracy versus 20.12 and 53.19 for a handcrafted-feature MLP.

Go to primary source ↓
e-detectionPDF p. 6, Table 4; p. 7, Section 5.1, detection paragraphInspect

Connected components of segmentation outputs supply object proposals; details are deferred to supplementary material. Mean AP at 3D IoU 0.5 is 24.24 versus 18.22 for Armeni et al.; sofa AP is lower, 4.76 versus 6.78.

Go to primary source ↓
e-poolingPDF p. 7, Figure 5, embedded accuracy table, caption and Section 5.2Inspect

The visually inspected accuracy table reports unsorted MLP 24.2, sorted MLP 45.0, LSTM 78.5, attention sum 83.0, average pooling 83.8 and max pooling 87.1 on the ModelNet40 design-analysis task.

Go to primary source ↓
e-transformsPDF p. 7, Table 5 and Effectiveness of Input and Feature TransformationsInspect

Overall ModelNet40 accuracy is 87.1 without transforms, 87.9 with input transform, 86.9 with feature transform, 87.4 with regularized feature transform, and 89.2 with both. The adjacent text describes the best configuration as both transforms plus regularization.

Go to primary source ↓
e-robustnessPDF pp. 7–8, Section 5.2, Robustness Test and Figure 6 with captionInspect

The tests use the Figure 5 max-pooling architecture and unit-sphere normalization. The prose reports 2.4 and 3.8 accuracy drops at 50% missing data for furthest and random sampling. Figure 6 separately plots deletion, uniform-sphere outliers and independent Gaussian perturbations. Outlier robustness is conditional on seeing outliers during training.

Go to primary source ↓
e-criticalPDF p. 8, Section 5.3 and Figure 7; p. 5, Theorem 2Inspect

Rows show original shapes, critical points and upper-bound shapes, with colors encoding depth. The upper-bound construction admits points whose every point-function value is no greater than the corresponding global descriptor component. This is a condition on set membership, not geometric interpolation.

Go to primary source ↓
e-efficiencyPDF p. 8, Section 5.4 and Table 6 including captionInspect

Full PointNet has 3.5M parameters and 440M FLOPs/sample, vanilla 0.8M and 148M, Subvolume 16.6M and 3633M, and MVCNN 60.0M and 62057M. Baselines pool multiple rotations or views. Runtime prose names TensorFlow and a “1080X GPU” but gives no software version or full timing protocol.

Go to primary source ↓
e-supplementPDF p. 4, Joint Alignment Network; p. 5, Theorem 1 and application footnote; p. 6, segmentation architecture paragraph; p. 7, detection paragraphInspect

The main paper explicitly refers T-net details, a theorem proof, additional applications, modified segmentation and 3DCNN architectures, and detection details to separate supplementary material.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.