Source-linked AI summary
OctAttention: Octree-Based Large-Scale Contexts Model for Point Cloud Compression
Chunyang Fu, Ge Li, Rui Song, Wei Gao, Shan Liu
TL;DR
Voxel-based contexts become limited for sparse point clouds, motivating OctAttention's octree-based large-scale context model. It combines sibling and ancestor contexts with attention and masking, achieving state-of-the-art results on LiDAR and object point clouds while substantially reducing coding time.
Problem
Voxel-based context models have reduced context when handling sparse point clouds, limiting effective distribution modeling.
Method
OctAttention represents point clouds with octrees and uses tree-structured attention over sibling and ancestor contexts, with masking for parallel encoding.
Results
OctAttention achieves state-of-the-art compression on LiDAR and object point cloud datasets, including a 25.4% bitrate saving on SemanticKITTI versus G-PCC.
Takeaways & Limitations
The model supports a generic compression approach for both LiDAR and object point clouds and substantially reduces encoding and decoding time versus VoxelDNN.
Abstract
from arXiv · showhide
In point cloud compression, sufficient contexts are significant for modeling the point cloud distribution. However, the contexts gathered by the previous voxel-based methods decrease when handling sparse point clouds. To address this problem, we propose a multiple-contexts deep learning framework called OctAttention employing the octree structure, a memory-efficient representation for point clouds. Our approach encodes octree symbol sequences in a lossless way by gathering the information of sibling and ancestor nodes. Expressly, we first represent point clouds with octree to reduce spatial redundancy, which is robust for point clouds with different resolutions. We then design a conditional entropy model with a large receptive field that models the sibling and ancestor contexts to exploit the strong dependency among the neighboring nodes and employ an attention mechanism to emphasize the correlated nodes in the context. Furthermore, we introduce a mask operation during training and testing to make a trade-off between encoding time and performance. Compared to the previous state-of-the-art works, our approach obtains a 10%-35% BD-Rate gain on the LiDAR benchmark (e.g. SemanticKITTI) and object point cloud dataset (e.g. MPEG 8i, MVUB), and saves 95% coding time compared to the voxel-based baseline. The code is available at https://github.com/zb12138/OctAttention.
Introduction
OctAttention targets limitations of voxel-based contexts for sparse point clouds by using octrees, large-scale sibling and ancestor contexts, tree-structured attention, and masking. It reports state-of-the-art compression across LiDAR and object point clouds.
- Motivation: OctAttention encodes point clouds as octrees because voxels are inefficient for sparse point clouds.The octree context includes current-node ancestors, sibling nodes, and ancestors of sibling nodes.
- Evaluation: OctAttention is evaluated against state-of-the-art methods on SemanticKITTI, MVUB, and MPEG 8i.The experiments cover both LiDAR and object point cloud compression.
- Method: The tree-structured attention mechanism models dependencies among sibling nodes and their ancestors in an expanded receptive field.It emphasizes significant context nodes while filtering redundant or irrelevant information.
- Method: Masking enables parallel encoding of multiple octree nodes despite sibling contexts disabling prior parallelization strategies.The operation is intended to alleviate the computational drawback introduced by adding siblings to the context.
- Results: The generic geometry compression model achieves state-of-the-art performance on several large-scale LiDAR and object point cloud datasets.The contribution statement covers both point cloud categories without restricting the claim to a single dataset.
Related Work
Point cloud compression methods use tree structures to reduce geometric redundancy and voxel representations to preserve geometric patterns, while octree-based learning models increasingly capture node dependencies through learned contexts.
- Voxel-based methods classify voxel occupancy with neural networks and outperform G-PCC across lossless, lossy, and progressive geometric compression settings.
- Voxel representations naturally preserve geometric patterns compared with octrees, motivating their use despite differing structural properties.
- Tree structures reduce geometric redundancy by merging common point-cloud coordinates, including octrees, quadtrees, KD trees, prediction trees, and related variants.
- Earlier approaches modeled contexts with hand-crafted features, whereas OctSqueeze introduced octree-based deep entropy modeling of dependencies between nodes and multiple ancestors.
- OctAttention represents point clouds as octrees and embeds each node’s occupancy code, level, and octant within a context window containing sibling and ancestor information.
Methodology
OctAttention represents point clouds with octrees and predicts node occupancy from expanded sibling and ancestor contexts. Tree-structured attention selects useful context information, while masking reduces the encoding-time cost of the expanded receptive field.
- Context Model: The expanded context combines preceding sibling nodes with each node’s ancestor features to model dependencies across the octree.The context window is constructed during breadth-first traversal, and each node feature includes occupancy, level, and octant embeddings.
- Context Model: The model embeds occupancy, level, and octant information into fixed-length vectors before concatenating node and ancestor features.These embeddings expose low-level geometric information hidden in octree nodes and normalize variables with different scales.
- Tree-Structured Attention: Multi-head self-attention filters irrelevant nodes and weights similar sibling and ancestor features before the model outputs occupancy probabilities.The attention layers and multilayer perceptron produce a 255-dimensional probability for each node.
- Octree Structure: OctAttention encodes point clouds as octrees, whose occupancy codes are predicted sequentially and coded losslessly with an arithmetic coder.Only nonempty child cubes are subdivided, and the octree depth controls the representation resolution.
- Mask Operation: The mask assigns varied-length receptive fields so octree encoding can be partially parallelized despite sibling-aware prediction.With masking, receptive fields are 4, 5, and 6, and encoding time is reduced by 3 times relative to the unmasked case.
- Learning: The training objective optimizes cross-entropy between predicted occupancy codes and ground-truth codes.The estimated conditional probability guides arithmetic coding of the octree sequence.
Experiments
Experiments evaluate OctAttention on sparse LiDAR and object point cloud datasets. The selected datasets cover self-driving scans and dynamic or complete human-body point clouds at multiple precisions.
- LiDAR Dataset: SemanticKITTI provides 43,552 sparse LiDAR scans containing 4,549 million points for self-driving evaluation.Sequences 00–10 are used for training and sequences 11–21 for testing.
- Object Point Cloud Dataset: MVUB and MPEG 8i provide dynamic half-body and complete human-shape point clouds with 9–12-bit precision.Testing sequences use resolutions and point clouds not seen during training.
Experimental Details
The experiments compare OctAttention under established LiDAR and object-point-cloud settings, using standard reconstruction metrics and multiple baselines. Training and testing configurations vary octree depth, quantization, and geometry precision to assess bitrate control and robustness.
- LiDAR Compression: LiDAR lossy compression varies octree depth from 8 to 12 and compares OctAttention with VoxelContext-Net and OctSqueeze.The comparison follows VoxelContext-Net’s training and testing setting because the baseline source codes are unavailable.
- Object Point Cloud Compression: Object compression uses lossless settings and compares OctAttention with P(full), VoxelDNN, MSVoxelDNN, and common hand-crafted methods.Models are tested at different octree depths and geometry precisions.
- Training and Testing: A single depth-12 model handles all LiDAR layers, while the object model is trained on depths 9 and 10 and evaluated across geometry precisions.LiDAR testing truncates the octree over depths 8–12 for different bitrates.
- Evaluation Metrics: Evaluation uses D1 PSNR and D2 PSNR for lossy geometry reconstruction, along with chamfer distance and PSNR peak values.D1 and D2 are standard MPEG geometry-quality metrics calculated with the pc error tool.
Experiment Results
OctAttention improves compression across LiDAR and object point cloud benchmarks, outperforming the compared baselines at evaluated bitrates and achieving substantial bitrate savings.
- OctAttention outperforms VoxelContext-Net and other baselines at all SemanticKITTI bitrates.
- 25.4% bitrate savings on SemanticKITTI versus G-PCC exceed OctSqueeze’s less-than-4% savings.
- More than 11% relative bitrate reduction versus VoxelContext-Net occurs at high bitrates.
- 32.8% average gain over G-PCC is achieved for lossless object point cloud compression.
- The evaluation covers SemanticKITTI, MVUB, and MPEG 8i point cloud datasets.
Ablation Study and Analysis
Ablations associate larger contexts, attention, sibling features, and octree-based processing with improved performance, while masking supports efficient encoding and stable runtime.
- 14% bitrate savings result when the context window expands from 8 to 1024.
- The mask operation reduces I/O by N0 times and decreases encoding time as context size increases.
- Attention visualization highlights geometry-similar points, including lines, planes, surfaces, and curvature, within each context window.
- Removing attention or sibling features reduces the effectiveness of the model in the ablation study.
- OctAttention remains stable across varying point densities, whereas VoxelDNN performance drops on sparse point clouds.
- 95% encoding-time and 91% decoding-time savings versus VoxelDNN are reported for object point clouds.
Conclusion
The paper concludes that OctAttention is an octree-based entropy model using large-scale contexts for sparse and dense point cloud geometry compression, with state-of-the-art results on both dataset types.
- OctAttention exploits large-scale contexts in an octree-based entropy model for sparse and dense point cloud geometry compression.
- The model extends context with sibling nodes, applies attention to emphasize significant nodes, and uses masking for parallel encoding.
- State-of-the-art performance is reported on both LiDAR and object point cloud datasets.
1. Compression performance at high bitrates
At high bitrates, OctAttention maintains strong compression performance on point clouds tested across octree depths beyond its main training depth and compares favorably with OctSqueeze and MuSCLE.
- 19.4% average bitrate savings over OctSqueeze are reported across tested depths 11–16.
- 13.5% bitrate savings versus OctSqueeze and 8.5% versus MuSCLE occur at the highest bitrate.
- The model trained on depth-12 octrees is tested on depth-11–16 data, much of which is nearly absent from training.
- OctAttention achieves comparable low-bitrate and better high-bitrate performance than MuSCLE without inter-frame information.
2. Visualization of embedding
The paper visualizes embedded and original octree features after dimensionality reduction to assess whether embedding captures local geometry. Embedded features produce more geometrically coherent color patterns than original features.
- Feature construction: Embedded features combine a node’s occupancy, level, octant, and ancestor information through learned embedding representations.The original feature concatenates these inputs, while the embedded feature applies learned transformations to them.
- Visualization procedure: Principal Component Analysis reduces each embedded feature to three dimensions, whose vectors determine the visualization colors.The same visualization compares embedded features with original features without embedding.
- Observed features: Embedding describes local geometry patterns including surfaces, orientation, and curvatures.The visualization is conducted on ModelNet40 point clouds.
- Observed features: Original features may assign varying colors within the same surface or parallel planes because occupancy codes do not consistently reflect geometry.This issue appears across examples including tents, cars, beds, radios, airplanes, sinks, bookshelves, and stairs.
- Observed features: Embedded features render similar geometry patterns with similar colors, distinguishing corresponding upper, lower, left, and right surfaces.The paper reports blue and green colors for upper and lower surfaces, and yellow and red for left and right sides.
- Conclusion: The paper concludes that embedding provides an enhanced local geometry descriptor for modeling the point cloud geometry distribution.This connects the visualization to the entropy-modeling objective.