Abstract
LiDAR sensors are critical for autonomous systems but generate massive data volumes that challenge storage and real-time transmission. We present FLaTEC (Frequency-Latent Triplane Efficient Compression), a novel framework that projects 3D point clouds into 2D "triplane" representations. By disentangling high-frequency details from low-frequency structures, FLaTEC achieves state-of-the-art reconstruction quality with up to 94% better bitrate efficiency than standard codecs while maintaining real-time processing speeds (>16 FPS).
Introduction
LiDAR data presents a "triple threat" for compression: cubic complexity of voxel grids ($O(N^3)$), extreme sparsity in outdoor scenes, and the limitations of isotropic bitrate allocation in traditional deep learning codecs. Standard methods often treat all signal frequencies equally, which leads to "detail loss" where critical high-frequency features, such as vehicle boundaries and fine textures, become blurred or distorted.
Methodology
FLaTEC addresses these challenges through three core innovations:
- Latent Triplane Representation: Instead of processing a 3D voxel grid, FLaTEC projects 3D features onto three orthogonal 2D planes (Top, Front, and Side views). This reduces computational complexity to quadratic $O(N^2)$, significantly lowering the memory footprint.
- Frequency-Disentangled Encoder/Decoder: The model uses Frequency Decomposition (FD) blocks to isolate low-frequency "global" content from high-frequency "detail" priors.
- Local Spectrum-attention (LSA) Enhancer: To restore 3D connectivity lost during 2D projection, the authors utilize a frequency-domain attention mechanism, replacing expensive spatial 3D convolutions with efficient operations in the frequency domain.
As illustrated in Figure 6, traditional isotropic approaches treat all signal frequencies equally, often degrading key features. FLaTEC’s anisotropic approach allows the system to allocate bitrates more intelligently—preserving the structural "bones" of the scene even at high compression ratios by treating high-frequency details as compact priors.
Key Results and Evaluation
The model was evaluated on major large-scale outdoor LiDAR benchmarks, including SemanticKITTI, the Ford Dataset, and 40mILEN.
- Efficiency: FLaTEC outperforms the industry-standard G-PCC codec by 78% on SemanticKITTI and 94% on the 40mILEN dataset in terms of BD-rate.
- Speed: The system supports real-time encoding and decoding at over 16 FPS on a consumer GPU, well above the standard 10 Hz capture rate of most LiDAR sensors.
Figure 5 demonstrates the visual and quantitative superiority of FLaTEC. While G-PCC produces a 0.90M binary, FLaTEC achieves a comparable or better reconstruction with a file size of only 0.038M. Qualitative results show that FLaTEC preserves sharp boundaries for vehicles and buildings where other methods produce "ghosting" or blurred artifacts.
Conclusion
FLaTEC provides a scalable, real-time solution for LiDAR compression by leveraging the efficiency of 2D triplanes and the precision of frequency disentanglement. Future work will focus on reducing voxelization dependency and integrating compression with downstream tasks like 3D object detection and semantic segmentation.
Summary updated