Adaptive Hierarchical Down-Sampling for Point Cloud Classification

1. Introduction

• Down-sampling can reduce computation as large point clouds pass through a network, particularly with high-density sensors such as Velodyne-64.Source: §1 · p. 1
• Neighborhood-based methods incur the cost of k-NN search, while one learned sampling approach may require post-processing because its output is not guaranteed to be an input subset.Source: §1 · p. 2
• A fast, deterministic, content-sensitive down-sampling method is needed to integrate point reduction into deep networks, and CPL is introduced to meet that need.Source: §1 · p. 2
• CPL selects output points from the input and adaptively preserves critical points, while remaining deterministic, permutation-invariant, and global rather than neighborhood-based.Its global design avoids local nearest-neighbor search.Source: §1 · p. 2

2.1. Deep Learning on Point Clouds

• Voxel methods face rapidly growing network size and computation at higher resolution, while lower resolution increases quantization error; point-based networks avoid conversion-related resolution loss.Octree and kd-tree approaches still subdivide a bounding volume rather than exploiting local point geometry.Source: §2.1 · p. 2

2.2. CNNs on Point Cloud as Graphs

• Local spatial graph filtering uses patches to apply convolution-like operators, but only a few such methods are used for point-cloud graphs.Examples include MoNet, GCNN [7], and DCNN as variants, and KCNet, FoldingNet, ECC, and DGCNN [26] for point-cloud graphs.Source: §2.2 · pp. 2–3
• Spectral graph methods can lack spatial locality, and filters learned in the spectral domain may not generalize across varying graph structures.Source: §2.2 · pp. 2–3

2.3. Point Cloud Down-Sampling in Deep Networks

• Down-sampling is desirable because it can reduce later convolution costs and runtime memory, and may improve robustness to some input perturbations.Source: §2.3 · p. 3
• PointNet and DGCNN keep point counts fixed, while PointNet++ uses costly farthest-point sampling with overlapping k-NN partitions.Source: §2.3 · p. 3
• Graph max-pooling methods can select or generate less relevant points and provide no guarantee that important points reach downstream layers.KCNet and FoldingNet pool maximum features over neighborhoods using a pre-built k-NN graph.Source: §2.3 · p. 3
• Some methods use static spatial sampling or feature-space distances, while learned sampling in generates points not guaranteed to be input points, complicating contribution tracking.Source: §2.3 · p. 3
• CPL adaptively filters points by their learned-feature contribution to global max-pooling, avoiding local nearest-neighbor search.Source: §2.3 · p. 3

3. Proposed Solution

• The proposed layers are presented as a systematic way to down-sample point-cloud features in arbitrary neural-network architectures, rather than sampling a specific dataset.The section introduces CPL and WCPL as permutation-invariant layers for unordered point clouds.Source: §3 · p. 4

3.1. Critical Points Layer (CPL)

• CPL defines critical points as those contributing any feature to the global max-pooled vector, then gathers their input points for down-sampling.It deduplicates the contributing indices, sorts them by accumulated feature values, resizes the index vector to a fixed size, and gathers the corresponding points and features.Source: §3.1 · pp. 4–5
• Sorting feature values and resizing the sorted indices make CPL invariant to input-point permutation.The layer sorts by feature values rather than point indices, and nearest-neighbor resizing preserves the stated swap relationship.Source: §3.1 · pp. 5–6

3.2. Weighted Critical Points Layer (WCPL)

• WCPL weights points by how often they contribute to the maximum feature vector, increasing the sampling probability of more frequent contributors.It repeats point indices according to their contribution frequency, which lowers the probability of missing a critical point compared with CPL.Source: §3.2 · p. 6

3.3. Critical Points Net (CP-Net)

• CPL/WCPL blocks reduce point counts hierarchically after graph convolution, and CP-Net can use graph convolution methods such as EdgeConv, MoNet, GCNN, or DCNN.At each block, the point count is reduced by its down-sampling factor; previous-layer features for the same points may also be used or concatenated.Source: §3.3 · pp. 6–7

3.4. CP-Net for 3D Object Classification

• For 3D object classification, CP-Net takes an unordered n × 3 point cloud, applies spatial transformation and EdgeConv filtering, and uses CPL/WCPL to reduce points by factor 1/4, from n to n/4.Source: §3.4 · pp. 7–8

4.1. Data preprocessing

• ModelNet40 experiments use 9,843 training and 2,468 test models, each represented by 1,024 uniformly sampled points normalized to the unit sphere.Training data are augmented with random scaling, rotation, and shifting.Source: §4.1 · p. 8

4.2. Training Details

• Training uses Adam for 400 epochs; a batch of 32 on one Nvidia P100 takes 9–10 hours.The initial learning rate is 0.001 and decays by 0.5 every 200,000 steps.Source: §4.2 · p. 8

4.3. Statistical Results

• CP-Net/WCP-Net rank behind but surpass the other methods in Table 1 and the ModelNet40 benchmark leaderboard.Table 1 reports classification accuracy for 1024 × 3 inputs.Source: §4.3 · p. 8, Table 1 · p. 9

4.4. Qualitative Results

• After 300 epochs, CPL retains classification-relevant parts even at 1/16 sampling, including lamp columns, airplane wings, and laptop corners.At 1/16, the point clouds contain 64 points; early training partially loses parts such as lamp columns and laptop screens.Source: §4.4 · pp. 8–10

4.5. Ablation studies

• Triple-kernel EdgeConv is more computationally complex than the double-kernel version, while CP-Net/WCP-Net outperform DGCNN in accuracy and use less computation.The comparison examines overall accuracy and execution time.Source: §4.5 · p. 10, Table 2 · p. 10
• Increasing bottleneck dimension improves accuracy but nearly saturates around 1,024 features; CP-Net reaches 89.35% with 64 features versus PointNet’s 89.20% with 1,024.Table 3 reports bottleneck-dimension effects on accuracy.Source: §4.5 · p. 10, Table 3 · p. 10
• Reducing from 1,024 to 64 points yields only a 0.73% accuracy difference; at a 1/4 ratio, random sampling scores 91.47% versus 92.33% for CPL and 92.41% for WCPL.Random sampling is viable in training, but at inference it can produce different indices and nondeterministic classifications for the same input.Source: §4.5 · pp. 10–11, Table 4 · p. 11
• CP-Net/WCP-Net have the largest models, reducible by lowering bottleneck dimension; they run faster than DGCNN and PointNet++ but slower than PointNet and KCNet.The execution-time comparison uses a single P100 GPU and Intel 1.2 GHz CPU with batches of 32 1,024-point clouds.Source: §4.5 · p. 11

• CP-Net combines CPL and WCPL in a hierarchical system, and experiments report competitive classification accuracy against previous state-of-the-art methods.Source: §5 · p. 11
• CPL dynamically down-samples unordered data, while the proposed approach is designed to pass important points to later network layers.Source: §5 · p. 11
• Table 5 reports model size, P100 GPU inference time, and classification accuracy, but its caption provides no values.Source: Table 5 · p. 11
• The authors identify auto-encoders for unordered data such as point clouds as a possible future application under investigation.Source: §5 · p. 12