Point-SkipNet: Lightweight 3D Classifier

• The paper’s main contribution is a lightweight 3D point cloud classifier that achieves high accuracy with only 1.47M parameters using efficient sampling and concatenation-based skip connections.
• Point-SkipNet leverages static radius-based grouping and shared MLP feature extraction to capture local geometric details while preserving spatial context through skip concatenation.
• The paired ModelNet-R benchmark refines labeling and removes artifacts, highlighting how improved data curation enhances model performance and reliability.

Searching arXiv for the specified paper to ground the article in the cited source. Point-SkipNet is a lightweight, neighborhood-based graph model for 3D point cloud classification introduced alongside ModelNet-R, a refined version of ModelNet40 intended to address inconsistent labeling, 2D artifacts, size mismatches, and inadequate class differentiation (Saeid et al., 5 Sep 2025). The model is designed around efficient sampling, radius-based neighborhood grouping, shared MLP feature extraction, symmetric max aggregation, and skip connections, with the stated goal of achieving high classification accuracy under a small parameter budget. In the paired formulation presented in the source paper, architectural efficiency and benchmark quality are treated as coupled factors: Point-SkipNet supplies a compact classification pipeline, while ModelNet-R supplies a cleaner evaluation and training substrate for measuring geometric discrimination more reliably (Saeid et al., 5 Sep 2025).

1. Position within 3D point cloud classification

The source paper situates Point-SkipNet against two constraints in 3D point cloud classification: the high computational cost of existing high-accuracy point models, especially MLP-heavy or transformer-based systems, and the need for robust local geometric encoding without dynamic-graph construction overhead or attention costs (Saeid et al., 5 Sep 2025). It further argues that stable gradient flow and retention of spatial detail are important in hierarchical pipelines, motivating the use of skip connections to avoid over-compression.

The problem setting assumes input point clouds represented as $p_i = (x_i, y_i, z_i)$ and $P = \{p_i\}_{i=1}^N$, with classification expressed as $C = CLS(P)$ (Saeid et al., 5 Sep 2025). Training uses augmented point sets, written as $P_{aug} = Aug(P)$, with rotations and other transforms applied to improve invariance (Saeid et al., 5 Sep 2025). The paper does not fix $N$ numerically, and it likewise does not specify exact neighborhood cardinalities or stage widths in the textual description.

A central claim of the work is that classification accuracy and model efficiency are constrained not only by architecture choices but also by dataset quality. This leads to a paired contribution structure: a refined benchmark, ModelNet-R, and a compact classifier, Point-SkipNet (Saeid et al., 5 Sep 2025). This suggests that Point-SkipNet should be understood not merely as an isolated architecture, but as part of an argument about how benchmark curation affects the apparent performance frontier for lightweight point models.

2. ModelNet-R as the benchmark context

ModelNet-R is introduced as a meticulously refined version of ModelNet40, motivated by four documented issues in the original benchmark: inconsistent labeling, 2D artifacts, size mismatches, and inadequate class differentiation (Saeid et al., 5 Sep 2025). The source attributes these issues to mislabeled samples, multi-object scenes, flat or near-planar shapes lacking volumetric depth, normalization effects that collapse real-world scale differences, and ambiguous class definitions such as plant versus flower_pot versus vase or cup versus bowl.

The refinement process consists of manual visual inspection and expert cross-referencing to identify mislabeled or ambiguous samples; correction when clearly identifiable; removal otherwise; elimination of 2D or flattened instances lacking volumetric depth; clarification of class boundaries; and filtering of multi-object scenes, unrecognizable shapes, and mislabels identified in confusion analyses and manual review (Saeid et al., 5 Sep 2025). The class clarifications are explicit: plant includes only plant samples; flower_pot denotes plant plus pot together; vase denotes an empty pot with no plant; cup must include a handle; and bowl corresponds to wide, low-height hemispherical shapes.

The paper states that the original ModelNet contains 40 classes and 12,311 instances, and that ModelNet-R refines 5 classes: flower_pot, plant, vase, cup, and bowl (Saeid et al., 5 Sep 2025). The number of classes remains 40, while no new global instance count across all classes is provided after refinement, and no train/test split changes are explicitly reported. The text implies that the standard split is preserved, with refined labels and filtered samples within the affected classes.

Class	Modification summary	Final count or retained count
Flower_pot	91 moved from other categories; 72 moved to vase; 5 to bowl; 1 removed	Final 262
Plant	171 moved to flower_pot; 16 removed	152 retained
Vase	55 moved out to cup; 2 moved to bowl; 2 removed	571 retained
Cup	55 moved from vase; 1 moved to bowl	43 retained
Bowl	Inflow/outflow as above	Total 84 before moves

The paper emphasizes that common normalization can obscure true scale and that ModelNet-R therefore emphasizes clearer geometries and labeling to mitigate reliance on size signals (Saeid et al., 5 Sep 2025). Since specific normalization procedures beyond relabeling, removal, and class clarification are not exhaustively detailed, stronger claims about scale restoration would go beyond the source. A plausible implication is that the benchmark redesign is intended less to preserve metric scale directly than to reduce the confounding effect of artifacts and ambiguous semantics.

3. Architectural design of Point-SkipNet

Point-SkipNet is described as a multi-stage sampling-grouping-MLP-pooling architecture with concatenation-based skip connections and a final global descriptor followed by a small fully connected classifier head (Saeid et al., 5 Sep 2025). Its core operators are Farthest Point Sampling (FPS), static radius-based ball query grouping, shared MLP feature extraction, symmetric max pooling, and a skip mechanism that concatenates center features with pooled neighborhood features.

Sampling and grouping are defined as follows. Given an input set of points, FPS selects a subset $\{p_j\}_{j=1}^{N_j} = FPS(\{p_i\}_{i=1}^{N_i})$ with $N_j < N_i$ (Saeid et al., 5 Sep 2025). For each sampled center $p_j$, a local neighborhood is formed by ball query, $G_j = \{ p_i \mid ||p_i - p_j|| \le r \}$, and the grouped representation is $$G_P = \{G_j\}_{j=1}^{N_j} = ball\_query(\{p_i\}, \{p_j\}, r)$$, with shape $P = \{p_i\}_{i=1}^N$0, where $P = \{p_i\}_{i=1}^N$1 is the neighbor count and $P = \{p_i\}_{i=1}^N$2 the channel dimensionality (Saeid et al., 5 Sep 2025). If additional features $P = \{p_i\}_{i=1}^N$3 exist, they are gathered by indices through $P = \{p_i\}_{i=1}^N$4.

This construction induces a static, radius-based neighborhood graph centered on sampled points and avoids dynamic graph recomputation (Saeid et al., 5 Sep 2025). The source explicitly states that no explicit edge-feature message passing is used; instead, local patches are encoded by MLPs and symmetric aggregation. It also states that the design contains no dynamic graph updates, no attention weights, and no explicit edge-feature message passing of the form $P = \{p_i\}_{i=1}^N$5 (Saeid et al., 5 Sep 2025).

At each stage, the processing pipeline is given by:

1. sampling and grouping,
2. MLP feature extraction,
3. max pooling across grouped neighbors,
4. skip connection via concatenation with center features,
5. dimensionality reduction via a second MLP (Saeid et al., 5 Sep 2025).

The corresponding equations are $P = \{p_i\}_{i=1}^N$6, where $P = \{p_i\}_{i=1}^N$7 and $P = \{p_i\}_{i=1}^N$8; $P = \{p_i\}_{i=1}^N$9; $C = CLS(P)$0; and $C = CLS(P)$1, with $C = CLS(P)$2 (Saeid et al., 5 Sep 2025). The global representation is then $C = CLS(P)$3, and the classifier outputs $C = CLS(P)$4 (Saeid et al., 5 Sep 2025).

The paper identifies this skip design as non-residual and concatenation-based, preserving spatial/context channels alongside newly aggregated features before reduction (Saeid et al., 5 Sep 2025). It notes that a canonical residual form would be $C = CLS(P)$5, with $C = CLS(P)$6 the identity or a projection, but states that Point-SkipNet realizes the primary skip pathway via concatenation rather than additive identity in order to retain richer information from both sources before dimensionality reduction.

4. Skip connections, efficiency, and operator choices

The architectural identity of Point-SkipNet depends heavily on its skip formulation. The paper evaluates two skip modes—concatenation and addition—and reports that concatenation is superior (Saeid et al., 5 Sep 2025). The two alternatives are explicitly written as

$C = CLS(P)$7

and $C = CLS(P)$8, with the latter used in ablation and described as empirically inferior (Saeid et al., 5 Sep 2025).

The reported ablation values are 93.79 OA / 92.49 mAcc for concatenation and 93.64 OA / 92.09 mAcc for addition (Saeid et al., 5 Sep 2025). The accompanying interpretation is that concatenation-based skips provide multi-scale fusion by preserving raw spatial/context features alongside pooled local descriptors, preventing lossy compression across stages and easing optimization by improving gradient flow. Empirically, the paper argues that retaining both feature streams before reduction is beneficial (Saeid et al., 5 Sep 2025).

From an efficiency standpoint, Point-SkipNet has a parameter count of 1.47M (Saeid et al., 5 Sep 2025). The source attributes its qualitative efficiency to the avoidance of dynamic graph recomputation and attention blocks, and to the use of FPS plus ball-query grouping with shared MLPs and max pooling (Saeid et al., 5 Sep 2025). Memory usage and FLOPs are not reported. The paper also does not specify exact values for $C = CLS(P)$9, $P_{aug} = Aug(P)$0, $P_{aug} = Aug(P)$1, or per-layer widths, so claims about scaling behavior beyond the stated qualitative design would be inferential.

The omission of BatchNorm, LayerNorm, gating, and attention details is explicit: feature normalization, nonlinearity, and gating or attention are not specified in the text, and the design emphasis is on lightweight MLPs and pooling (Saeid et al., 5 Sep 2025). This makes the model definition relatively operator-centric rather than implementation-microarchitectural in the published description.

5. Training setup and reported experimental performance

Training uses cross-entropy classification loss, $P_{aug} = Aug(P)$2, where $P_{aug} = Aug(P)$3 is the one-hot label and $P_{aug} = Aug(P)$4 the softmax of scores (Saeid et al., 5 Sep 2025). The optimizer is Adam with initial learning rate 0.001, batch size 32, and 200 epochs (Saeid et al., 5 Sep 2025). Training and experiments were conducted on an NVIDIA GeForce RTX 3080 GPU. Inference latency and throughput are not reported.

The evaluation metrics are Overall Accuracy (OA) and Mean Class Accuracy (mAcc) (Saeid et al., 5 Sep 2025). A principal empirical finding is that training on ModelNet-R consistently improves OA and mAcc across multiple models. The reported comparisons are as follows.

Model	ModelNet $P_{aug} = Aug(P)$5 ModelNet-R	Parameters
PointNet	89.20 / 86.00 $P_{aug} = Aug(P)$6 91.39 / 88.79	3.47M
PointNet++ (SSG)	– / – $P_{aug} = Aug(P)$7 94.02 / 92.40	1.47M
PointNet++ (MSG)	90.70 / – $P_{aug} = Aug(P)$8 94.06 / 91.80	1.74M
Point-NN (Walk in the Cloud)	81.80 / – $P_{aug} = Aug(P)$9 84.75 / 77.65	0.00M
DGCNN	92.90 / 90.20 $N$0 94.03 / 92.64	1.80M
CurveNet	93.80 / – $N$1 94.12 / 92.65	2.04M
PointMLP	94.10 / 91.10 $N$2 95.33 / 94.30	12.60M
Point-SkipNet	92.29 / 89.84 $N$3 94.33 / 92.93	1.47M

Within this table, Point-SkipNet improves by +2.04 OA and +3.09 mAcc when moving from ModelNet to ModelNet-R (Saeid et al., 5 Sep 2025). The source characterizes this result as state-of-the-art accuracy on ModelNet-R with a substantially lower parameter count compared to contemporary models, and specifically states that Point-SkipNet achieves state-of-the-art accuracy among lightweight MLP/graph neighborhood methods in the refined setting with only 1.47M parameters (Saeid et al., 5 Sep 2025). Because PointMLP reaches 95.33 / 94.30 on ModelNet-R at 12.60M parameters, the claim is best read in the paper’s stated lightweight framing rather than as an unconditional best-overall-performance statement across all parameter scales.

The augmentation ablation reports OA/mAcc of 93.93 / 92.55 for rotation, 93.76 / 92.59 for anisotropic scaling, 93.75 / 92.49 for translation, 93.56 / 92.19 for jitter, 93.49 / 92.25 for all augmentations, and 93.72 / 92.56 for the main baseline augmentation set (Saeid et al., 5 Sep 2025). Rotation alone yields the highest OA, while “All augmentations” does not outperform rotation alone. This is presented as an empirical result rather than a general rule.

6. Interpretation, limitations, and possible extensions

The paper’s analysis attributes the effectiveness of Point-SkipNet partly to the interaction between hierarchical local encoding and skip connections (Saeid et al., 5 Sep 2025). The stated interpretation is that the local encoders capture fine geometry while skips ensure early-stage spatial cues are not discarded. In parallel, the paper argues that cleaned labels and removal of 2D artifacts reduce label noise and avoid pathological local structures that mislead encoders, and that clearer class boundaries improve discriminative learning, enabling lightweight architectures to reach high accuracy without heavy attention or deep capacity (Saeid et al., 5 Sep 2025).

The broader implication drawn in the source is that better data distributions decrease overfitting to spurious cues and sharpen decision boundaries, which is especially useful for small, efficient models (Saeid et al., 5 Sep 2025). This suggests that some of the performance usually attributed to larger architectures may in fact depend on the impurity or ambiguity of the benchmark itself. That inference is consistent with the documented gains across diverse baselines, although the paper does not quantify causal attribution separately for each source of refinement.

Several limitations are explicitly stated. First, the refinement currently covers 5 of 40 classes, so extending it to all classes is identified as a direction for further consistency improvements (Saeid et al., 5 Sep 2025). Second, common normalization can suppress size cues, and future techniques to restore useful scale information are suggested (Saeid et al., 5 Sep 2025). Third, Point-SkipNet is evaluated only on ModelNet and ModelNet-R; broader validation on real-world benchmarks such as ScanObjectNN and ShapeNet is listed as future work (Saeid et al., 5 Sep 2025). Fourth, the hyperparameters $N$4, $N$5, $N$6, and $N$7 are not specified, and the paper notes that performance could be sensitive to these choices.

Potential extensions named in the source include segmentation and detection by adapting the Sample-and-Group blocks to per-point labeling or proposal generation, dynamic graphs by refreshing neighborhoods per stage or by feature-space grouping if computational budget permits, and self-supervised pretraining using contrastive or reconstruction objectives on ModelNet-R to further boost accuracy under small parameter budgets (Saeid et al., 5 Sep 2025). These are prospective directions rather than demonstrated capabilities.

7. Implementation profile and reproducibility boundaries

The source provides a high-level pseudocode sketch for Point-SkipNet, beginning with input point cloud $N$8 and optional features, then augmentation, FPS subsampling, radius-based ball query around each sampled center, shared neighborhood MLP processing, neighbor-axis max pooling, concatenation with center features, MLP-based reduction, global max pooling over sampled centers, and final fully connected classification with $N$9 trained by cross-entropy loss (Saeid et al., 5 Sep 2025). This matches the operator-level architecture description and makes clear that the classifier is a global-shape model rather than a dense pointwise predictor.

The accompanying repository is described as providing dataset loaders and evaluation scripts implementing the sampling, grouping, MLP, and OA/mAcc reporting pipeline, with ablation toggles for augmentation and skip mode (Saeid et al., 5 Sep 2025). The reported training configuration matches the paper: Adam, learning rate $\{p_j\}_{j=1}^{N_j} = FPS(\{p_i\}_{i=1}^{N_i})$0, batch size 32, 200 epochs, on an RTX 3080 GPU. However, seeds and full reproducibility settings are not detailed in the text; the source notes that default PyTorch/cuDNN determinism and seeds may be used as in typical practice (Saeid et al., 5 Sep 2025). Because this is framed as a possibility rather than a documented configuration, it should not be treated as a definitive reproducibility prescription.

In aggregate, Point-SkipNet is best characterized as a compact neighborhood-MLP classifier whose main technical commitments are static radius-based grouping, symmetric max aggregation, and concatenation-based skip fusion, evaluated in tandem with a benchmark-cleaning effort that seeks to remove confounding artifacts from ModelNet40 (Saeid et al., 5 Sep 2025). Its significance in the source lies less in introducing an entirely new operator family than in showing that a small 1.47M-parameter model can reach strong accuracy when paired with a refined benchmark designed to better reflect genuine 3D geometric discrimination.