Source-linked AI summary
OctSqueeze: Octree-Structured Entropy Model for LiDAR Compression
Lila Huang, Shenlong Wang, Kelvin Wong, Jerry Liu, Raquel Urtasun
TL;DR
LiDAR compression must reduce the storage burden of large, sparse point clouds while preserving reconstruction quality. OctSqueeze encodes points with an octree and models its symbols using a tree-structured conditional entropy model; it reports 10-20% lower bitrate at matched reconstruction quality and better downstream performance than prior methods.
Problem
LiDAR point clouds create substantial storage challenges, while sparse, non-grid data make directly applying deep compression methods non-trivial.
Method
OctSqueeze represents quantized LiDAR points with an octree and uses ancestor and decoding-available context in a tree-structured deep entropy model to encode octree symbols.
Results
10-20% lower bitrate than the previous state-of-the-art is reported at the same reconstruction quality, with better downstream segmentation and detection performance using compressed representations.
Takeaways & Limitations
The compressed representations reduce bitrate while retaining reconstruction quality and achieving lower error on downstream tasks than prior state-of-the-art work.
Takeaways & Limitations
Sequential decoding makes child input context unavailable, so the entropy model aggregates node features with ancestors rather than children.
Abstract
from arXiv · showhide
We present a novel deep compression algorithm to reduce the memory footprint of LiDAR point clouds. Our method exploits the sparsity and structural redundancy between points to reduce the bitrate. Towards this goal, we first encode the LiDAR points into an octree, a data-efficient structure suitable for sparse point clouds. We then design a tree-structured conditional entropy model that models the probabilities of the octree symbols to encode the octree into a compact bitstream. We validate the effectiveness of our method over two large-scale datasets. The results demonstrate that our approach reduces the bitrate by 10-20% at the same reconstruction quality, compared to the previous state-of-the-art. Importantly, we also show that for the same bitrate, our approach outperforms other compression algorithms when performing downstream 3D segmentation and detection tasks using compressed representations. Our algorithm can be used to reduce the onboard and offboard storage of LiDAR points for applications such as self-driving cars, where a single vehicle captures 84 billion points per day
1. Introduction
LiDAR produces massive, sparse point-cloud data whose raw matrix representation leaves structural redundancy underused. OctSqueeze addresses this with an octree and tree-structured entropy model, then evaluates compression across datasets and related approaches.
- A single Velodyne HDL-64 sensor generates over 84 billion points per day, creating onboard, offboard storage, and real-time communication challenges.
- Raw float-precision n × 3 matrices do not exploit the structured geometry and redundant local patterns present in typical scenes.
- Deep compression for LiDAR must address sparse, non-grid structure while finding a memory-efficient representation and an entropy model that encodes it efficiently.
- OctSqueeze constructs an octree representation and applies a tree-structured conditional entropy model before entropy coding the serialized symbols.
- 10-20% lower bitrate is achieved than the previous state-of-the-art at the same reconstruction quality across the evaluated results.
- Prior compression approaches include KD-trees, octrees, voxel grids, range images, and deep models operating on voxel or projected representations.
3. Octree-Structured Entropy Model
OctSqueeze represents LiDAR point clouds with an octree and learns a tree-structured conditional entropy model to encode occupancy symbols compactly. The model combines available node context with ancestor information before entropy coding.
- Overview: The method targets lossy LiDAR compression by reducing storage while preserving reconstruction quality through an octree-structured deep entropy model.
- 3.1. Octree Structure: An octree recursively partitions space into equal octants, stores each node’s 8-bit child-occupancy symbol, and subdivides only non-empty cells.Its resolution increases with tree depth, while adaptive subdivision preserves sparsity.
- 3.1. Octree Structure: Serialization traverses the octree into occupancy-code and leaf-offset bytestreams, preserving the tree exactly; quantization is the only lossy procedure.
- 3.2. A Deep Entropy Model for Entropy Coding: Predicted occupancy probabilities are passed to an entropy coder, where better distribution estimates reduce the achievable bitrate through cross-entropy minimization.
- 3.2. A Deep Entropy Model for Entropy Coding: The entropy model estimates each occupancy symbol’s probability conditioned on ancestor symbols and decoding-time context such as location, octant, level, and parent occupancy.Ancestor conditioning uses coarser geometry, while context features encode scene priors available during sequential decoding.
- 3.2. A Deep Entropy Model for Entropy Coding: Each node first receives an independent contextual embedding, then progressive parent-feature aggregation incorporates ancestral information along the octree traversal path.The model aggregates parent features rather than child features because child context is unavailable during sequential decoding.
4. Experiments
Experiments evaluate OctSqueeze on NorthAmerica and KITTI for reconstruction quality, bitrate, and downstream perception, including segmentation and detection. Across these evaluations, the method generally outperforms the compared compression baselines, with lower bitrate at matched reconstruction quality and stronger downstream performance.
- Datasets: 500K NorthAmerica scans train the entropy model, while 472 evaluation snippets provide reconstruction-quality and annotated perception-task evaluation.The NorthAmerica data cover urban and highway driving scenes; no additional filtering or processing is applied to the training scans.
- Datasets: SemanticKITTI contains 21,351 scans and 4.5 billion points, enabling domain-transfer evaluation despite differences in scene layout and sensor configuration.The dataset also provides dense point-wise labels for 25 classes.
- Baselines: The comparisons include Draco, the MPEG anchor, and Deep Range, spanning octree-based, off-the-shelf, and range-image compression approaches.Deep Range converts each LiDAR scan into a 2.5D polar range image before compression.
- Metrics: Reconstruction quality is measured with distance- and occupancy-based metrics, including PSNR, IOU, and Chamfer distance.Occupancy IOU uses 0.2 × 0.2 × 0.1 meter voxels.
- Compression Results: 10-20% lower bitrate than Draco and the MPEG anchor is achieved at the same reconstruction quality on NorthAmerica.Because OctSqueeze and the MPEG anchor use the same octree structure, their reconstruction quality is the same, while the learned entropy model reduces bitrate.
- Compression Results: On KITTI, the NorthAmerica-trained model significantly outperforms competing methods, particularly at lower bitrates.Qualitative comparisons also report more faithful reconstructions than range-image compression at comparable bitrates and better reconstruction quality than Draco at the same bitrate.
- Ablation Studies: Adding context features and increasing ancestor aggregation consistently lowers the entropy of the encoding model.The evaluated context features include octree level, parent occupancy, octant index, and spatial location; aggregation levels range from K = 0 to 4.
- Downstream Perception Tasks: At 5 bits-per-point, OctSqueeze improves NorthAmerica semantic-segmentation performance by 5-10% over Draco and MPEG, while also outperforming baselines on detection and on both datasets overall.The detection gains are smaller because the detection model is already robust across a range of bitrates.
5. Conclusion
OctSqueeze compresses LiDAR point clouds with a deep tree-structured entropy model over octree representations, exploiting sparsity and structural redundancy. Across two large-scale datasets, it reduces bitrate at matched reconstruction quality and lowers downstream-task error relative to prior work.
- OctSqueeze uses a deep tree-structured entropy model on octree point-cloud representations.The model incorporates context information to reduce the entropy of intermediate nodes.
- The method exploits sparsity and structural redundancy between points to reduce overall bitrate.
- The method is validated on two large-scale datasets.
- OctSqueeze significantly reduces bitrate at the same reconstruction quality compared with competing algorithms.
- Compressed representations achieve lower error on downstream tasks than prior state-of-the-art work.
Supplementary Material – OctSqueeze: Octree-Structured Entropy Model for
The supplementary material is associated with the listed authors and the January 2021 arXiv version of OctSqueeze.
- The paper lists Lila Huang, Shenlong Wang, Kelvin Wong, Jerry Liu, and Raquel Urtasun as authors.
- The listed author affiliation email domain is uber.com.
- The document is identified as arXiv:2005.07178v2, dated 8 January 2021.
Abstract
The supplementary material adds experiments assessing OctSqueeze’s efficacy, runtime, and qualitative behavior. It includes comparisons with Draco on NorthAmerica and KITTI for reconstruction quality and downstream task performance.
- The supplementary material reports additional experiments to further validate the proposed method’s efficacy.
- The authors benchmark runtime and demonstrate real-time encoding of LiDAR point clouds.
- Qualitative results on NorthAmerica and KITTI compare OctSqueeze with Draco for reconstruction quality and downstream task performance.
1. Additional Ablation Studies
The ablations examine aggregation depth, parental-context aggregation, and input context features in the entropy model. They support K = 4 aggregation levels and hierarchical parental information as important design choices.
- Additional Ablation Studies: The supplementary analysis evaluates aggregation depth, parental-context aggregation, and input context features for the entropy model.
- Number of Aggregations: K = 5 aggregations perform worse than K = 4 in bitrate reduction, supporting K = 4 as the best choice for this architecture.
- Parental Context Aggregation: Removing parental context aggregation performs worse than the K = 4 model and only performs as well as the K = 0 model.
- Parental Context Aggregation: The hierarchical feature-aggregation gains are attributed to the aggregation scheme rather than increased model capacity alone.
- Input Context Features: Adding the node’s octree level, parent occupancy symbol, octant index, and spatial location progressively decreases bitrate.
2. Additional Baselines
The paper compares OctSqueeze with range-image and voxel-based compression baselines. Octree-based representations and entropy modeling provide stronger reconstruction-quality results than the tested alternatives.
- Range-image baselines: Deep Range outperforms JPEG Range across all reconstruction quality metrics on NorthAmerica and KITTI.
- Range-image baselines: OctSqueeze significantly outperforms Deep Range and JPEG Range, using an octree representation and octree-structured entropy model.
- Voxel baseline: 33.76 point-to-plane PSNR at 26.81 bitrate makes the voxel baseline substantially worse than Draco and OctSqueeze.Draco reaches 48.47 PSNR at 2.778 bpp, while OctSqueeze reaches 48.95 PSNR at 1.61 bpp.
- Voxel baseline: The voxel method’s underperformance indicates that dense voxel representations may not fit sparse, high-frequency LiDAR point clouds well.
3. Runtime
Runtime is reported by octree depth and separates octree construction, entropy-network execution, and range coding. The implementation achieves end-to-end encoding in real time on the tested workstation.
- Runtime breakdown: Octree construction, network execution, and range coding are reported separately in milliseconds by maximum octree depth.Table 4 uses K = 4 aggregations and defines the three runtime components.
- Runtime performance: OctSqueeze achieves end-to-end encoding in real time on an Intel Xeon E5-2687W CPU and Nvidia GeForce GTX 1080 GPU.Octree building and range coding use C++, while the entropy model uses Python with PyTorch.
4. Additional Qualitative Results
Qualitative evaluations compare OctSqueeze with Draco across reconstruction and downstream perception at multiple bitrates. OctSqueeze attains better results at comparable or lower bitrates.
- Evaluation scope: The qualitative figures compare OctSqueeze and Draco for reconstruction quality, semantic segmentation, and object detection across bitrates.
- Qualitative comparison: OctSqueeze can attain better results than Draco at comparable or even lower bitrates.
5. Change Log
The ArXiv v2 change log records corrected reconstruction-metric definitions and revised OctSqueeze decoding-runtime estimates.
- Metric corrections: ArXiv v2 corrects the definitions of symmetric point-to-point Chamfer distance and symmetric point-to-plane PSNR.
- Runtime correction: ArXiv v2 also corrects OctSqueeze’s decoding-runtime estimates.