Research map / Point clouds / Point cloud classification
TopNet: Transformer-Efficient Occupancy Prediction Network for Octree-Structured Point Cloud Geometry Compression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TopNet, a lightweight CNN-Transformer hybrid octree context model with four modules (LeCE, AL-SWA, SG-CM, LNOP) for entropy coding of point cloud geometry, reaching SOTA bitrate reduction with about 3.4M parameters.
Problem
Octree-based point cloud geometry compression needs context models that capture long-range dependencies without the parameter and compute cost of Transformers, while preserving fine-grained local detail.
Approach
Combine convolutions with Transformer attention: depth-wise/point-wise convolutions replace absolute positional encoding and linear projections, a learned query embedding and length-scaled cosine attention keep attention stable across window lengths, a spatial-gated channel mixer replaces the FFN, and a star-operation predictor head produces the occupancy distribution.
Key results
- On SemanticKITTI, TopNet achieves a 33.4% bitrate reduction over G-PCC versus 24.7% for OctAttention.
- On lossless human body PCGC, TopNet averages 43.4% BPP reduction vs G-PCC on 8iVFB and 35.4% on MVUB.
- TopNet uses 3.37M parameters versus 4.23M for OctAttention, and ablation shows the full model lowers BPP on 8iVFB from 0.626 to 0.558.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LeCE (ablation) | TopNet · 8iVFB | BPP: 0.626 -> 0.592; −5.43% | Helped |
| Add LeCE (ablation) | TopNet · MVUB | BPP: 0.745 -> 0.744; −0.13% | Helped |
| Add AL-SWA (ablation) | TopNet · 8iVFB | BPP: 0.592 -> 0.565; −9.74% | Helped |
| Add AL-SWA (ablation) | TopNet · MVUB | BPP: 0.744 -> 0.716; −3.89% | Helped |
| Add SG-CM (ablation) | TopNet · 8iVFB | BPP: 0.565 -> 0.562; −10.22% | Helped |
| Add SG-CM (ablation) | TopNet · MVUB | BPP: 0.716 -> 0.712; −4.43% | Helped |
| Add LNOP (ablation) | TopNet · 8iVFB | BPP: 0.562 -> 0.558; −10.86% | Helped |
| Add LNOP (ablation) | TopNet · MVUB | BPP: 0.712 -> 0.707; −5.10% | Helped |
| Use TopNet instead of G-PCC | 12-bit quantized SemanticKITTI · SemanticKITTI | bitrate reduction over G-PCC: 32.99% | Helped |
| Use length-scaled cosine attention instead of scaled cosine attention (ablation) | TopNet · 8iVFB, window length 8192 | BPP: 8192, 0.577 | Helped |
Surprising findings
- Encoding time increases slightly over OctAttention (0.38s vs 0.19s on LiDAR), although total runtime is reported as lower.
Limitations
- Encoding time is slightly higher than OctAttention.
- Baseline implementations were limited, so generalization comparisons on nuScenes, LiDAR-CS and ScanNet included only OctAttention and G-PCC.
Experimental setup
Datasets: SemanticKITTI 8iVFB MVUB nuScenes LiDAR-CS ScanNet
Benchmarks: SemanticKITTI nuScenes LiDAR-CS ScanNet 8iVFB MVUB
Metrics: bits per point (BPP) D1 PSNR Chamfer distance bitrate reduction over G-PCC parameters encoding/decoding time memory
Baselines: OctSqueeze VoxelContext-Net SparsePCGC OctAttention OctFormer EHEM ECM-OPCC G-PCC VoxelDNN MSVoxelDNN
Builds on: OctAttention OctFormer scaled cosine attention Inception-Residual Network star operation
Closest papers by meaning
- UniPCGC: Towards Practical Point Cloud Geometry Compression via an Efficient Unified Approach (AAAI 2025 · similarity 0.87)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.86)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.86)
- OctreeOcc: Efficient and Multi-Granularity Occupancy Prediction Using Octree Queries (NeurIPS 2024 · similarity 0.86)
- Unsupervised Occupancy Learning from Sparse Point Cloud (CVPR 2024 · similarity 0.85)
- Multi-Scale Neighborhood Occupancy Masked Autoencoder for Self-Supervised Learning in LiDAR Point Clouds (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.