LiDAR Point Cloud Feature Transmission

• LPC-FT is a communication regime that maps LiDAR point clouds to compact latent or semantic features instead of raw coordinates.
• It uses deep learning techniques such as DeepJSCC and hierarchical local-global encoding to optimize geometry and task fidelity over constrained channels.
• The approach supports applications like 3D object detection and collaborative perception by balancing robust reconstruction with effective feature transmission.

Searching arXiv for the cited LPC-FT and related papers to ground the article in current literature. LiDAR Point Cloud Feature Transmission (LPC-FT) denotes a communication regime in which a LiDAR point cloud is not sent as raw coordinates or solely as a conventional compressed geometry bitstream, but is first mapped into compact latent, semantic, or task-driven representations that are then transmitted over constrained links for reconstruction or downstream perception. In the literature, LPC-FT appears in several closely related forms: end-to-end semantic communication for geometry reconstruction, hierarchical local-global feature transmission, and cloud-edge feature coding for 3D object detection (Bian et al., 2023, Xie et al., 2024, Liu et al., 8 Sep 2025, Tian et al., 4 Mar 2026). The common problem setting is a point cloud such as

$$\mathcal{P} = (\mathbf{X}, \mathbf{F}), \quad \mathbf{X} = \{\mathbf{x}_i \in \mathbb{R}^3\}_{i=1}^N,$$

or, in geometry-only form,

$$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$

with the transmitted object being a feature tensor, latent vector, neural representation, or compacted detector feature rather than the original unordered point set (Bian et al., 2023, Xie et al., 2024).

1. Problem formulation and conceptual scope

LPC-FT is motivated by the fact that “3D point cloud is a three-dimensional data format generated by LiDARs and depth sensors” and is “increasingly used in autonomous driving, AR, robotics, etc.”, while being large scale, sparse, irregular, unordered, and produced under tight bandwidth and latency constraints (Bian et al., 2023). The broader wireless systems setting is multi-sensor uplink to an access point or edge server for “fusion of multi-view data, segmentation, detection, or other inference,” under limited bandwidth, volatile channels, and delay-sensitive applications (Shao et al., 2024).

Within this space, LPC-FT is narrower than generic point cloud transmission. The distinguishing property is that “the transmitted data are semantic/latent features of the point cloud, not raw coordinates,” or, in cloud-edge collaborative perception, compacted detector features extracted from the deepest network stage (Bian et al., 2023, Tian et al., 4 Mar 2026). By contrast, several adjacent works study efficient transmission of compressed raw geometry, such as roadside LiDAR compression, MPEG G-PCC/L3C2 transmission for remote segmentation, or lossless predictive coding; these are highly relevant baselines, but they are not feature transmission in the strict sense (Zimmer et al., 2024, Fernandes et al., 27 Sep 2025, Shibu et al., 24 Mar 2026).

A recurrent conceptual split in the literature is between source coding that optimizes geometric fidelity and source/channel designs that optimize task or semantic fidelity. In the former, the target is typically point-to-point or point-to-plane distortion. In the latter, the target is graceful degradation under SNR variation, stable reconstruction quality, or downstream detection and segmentation performance (Bian et al., 2023, Fernandes et al., 27 Sep 2025).

2. End-to-end semantic transmission of LiDAR geometry

A foundational LPC-FT formulation is the DeepJSCC-style geometry autoencoder introduced in “Over-the-Air Learning-based Geometry Point Cloud Transmission” (Bian et al., 2023). Its SEmantic Point cloud Transmission (SEPT) pipeline uses a point-cloud encoder $h(\cdot)$, power normalization, an AWGN channel, and a decoder $g(\cdot)$. The channel model is

$$\mathbf{y}' = \mathbf{z}' + \mathbf{w}, \quad \mathbf{w}\sim \mathcal{CN}(0,N_0),$$

and the bandwidth budget is expressed by the channel uses per point

$\text{CPP} = \frac{n}{2N}.$

In the LPC-FT interpretation, $h(\cdot)$ is the LiDAR feature extractor and JSCC encoder, $\tilde{\mathbf{z}}\in\mathbb{R}^n$ is the transmitted LiDAR feature vector, and $g(\cdot)$ is the feature decoder that reconstructs geometry from noisy features (Bian et al., 2023).

SEPT obtains its latent by iterative downsampling and feature extraction. The encoder uses Farthest Point Sampling, local neighborhood aggregation, and a Point Transformer self-attention module with vector attention

$$\mathbf{f}^{(2)}_i = \sum_{j\in \mathbb{N}(i)} \mathrm{Softmax}\big(\gamma(\phi(\mathbf{f}^{(1)}_i) - \phi(\mathbf{f}^{(1)}_j))+\delta\big)\odot (\alpha(\mathbf{f}^{(1)}_j) + \delta),$$

with

$$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$0

After two downsampling and attention stages, SEPT produces a downsized point cloud $$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$1, where $$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$2 are the learned semantic features. A central design decision is that $$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$3 are not transmitted; SEPT instead transmits only feature information and lets the decoder infer coordinates, thereby avoiding the digital overhead and “cliff/leveling effects” associated with separate source and channel coding (Bian et al., 2023).

The latent is further compacted by global max pooling. In the reported ablation, max pooling yields much higher D1 PSNR than a linear projection at the same channel budget, for example “34.33 vs 26.41 dB at 0 dB SNR.” The decoder reconstructs a coarse latent geometry, refines it with self-attention, and then performs offset-based upsampling: $$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$4 with $$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$5. The model is trained end-to-end over the channel with the Chamfer distance

$$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$6

This makes the transmitted feature vector explicitly geometry-aware in a probabilistic channel sense (Bian et al., 2023).

The same paper extends the feature-transmission idea toward representational transmission. OTA-NeRF “learns to encode the voxelized point cloud into a neural network” and “transmits the learned neural network weights in an analog fashion along with few hyperparameters that are transmitted digitally,” while OTA-MetaNeRF addresses “the poor generality of the OTA-NeRF schemes” by encoding “different input point clouds into the latent vectors with shared neural network weights” (Bian et al., 2023). These variants show that, within LPC-FT, the transmitted object need not be a single latent vector; it may also be a neural representation.

3. Local-global semantic representations and hierarchical transmission

A second major LPC-FT pattern is the explicit decomposition of point-cloud information into local and global semantics. “Semantic Communication for Efficient Point Cloud Transmission” proposes a five-component architecture consisting of “local semantic encoder, global semantic encoder, channel encoder, channel decoder, and semantic decoder,” with two auxiliary modules, the patch extractor and the projection extractor (Xie et al., 2024).

The local path partitions the point cloud into overlapping patches. Using FPS to select $$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$7 centroids and kNN to gather $$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$8 neighbors, the method imposes the constraint

$$\mathbf{P} \in \mathbb{R}^{N \times 3}, \quad \mathbf{P} = \{(x_i, y_i, z_i)\}_{i=1}^N,$$9

so each original point appears in roughly two patches. Each patch is processed by a “PointNet++ + PointNet” pipeline to produce a local semantic matrix

$h(\cdot)$0

The global path generates multi-view projections with “MVTN-style learned views,” then encodes them into a low-dimensional global feature

$h(\cdot)$1

In the reported experiments, “very low-dimensional global codes (4–6 floats) already significantly improve reconstruction,” and “optimal D2 PSNR” occurs around $h(\cdot)$2 (Xie et al., 2024).

Transmission is hierarchical. The centroids $h(\cdot)$3 and global feature $h(\cdot)$4 are sent “losslessly” using “a strong channel code (e.g., PBRL-LDPC + HARQ),” whereas the local semantic feature $h(\cdot)$5 is sent “lossily via JSCC neural encoder over AWGN” (Xie et al., 2024). This division is technically significant because it assigns strong protection to low-rate global structure and lighter protection to high-rate local detail. At the receiver, the semantic decoder broadcasts $h(\cdot)$6 to all patches, concatenates it with the noisy decoded $h(\cdot)$7, and reconstructs patches by a lightweight upsampler: $h(\cdot)$8

Training is split into two stages. In Stage 1, semantic encoders and decoder are pretrained without channel noise. In Stage 2, the channel encoder and decoder are inserted, the semantic encoders are fixed, and the channel-facing modules are trained under AWGN using the same Chamfer loss. This “two-stage training” improves robustness, with the reported D2 PSNR comparison showing “47.705 vs 47.677” at 10 dB, “46.845 vs 46.078” at 5 dB, and “45.474 vs 43.788” at 0 dB (Xie et al., 2024). The same work reports that the method “maintain[s] a reconstruction quality of over 37dB even with severe channel noise” and that, under similar BPP, it achieves “4–6 dB higher PSNR than DPCC” (Xie et al., 2024).

This local-global pattern is one of the clearest architectural formulations of LPC-FT. It separates high-rate local geometry from low-rate global scene priors, uses different physical-layer protection levels for each, and reconstructs point sets from semantic codes rather than from raw octree or coordinate streams.

4. Collaborative perception and cloud-edge feature coding

The term LPC-FT is made explicit in “Synesthesia of Machines (SoM)-Aided LiDAR Point Cloud Transmission for Collaborative Perception,” which proposes “an efficient, robust and applicable LiDAR point cloud transmission system via the Synesthesia of Machines (SoM), termed LiDAR Point Cloud Feature Transmission (LPC-FT)” (Liu et al., 8 Sep 2025). In this system, the transmitter sends a compressed, density-preserving point-set representation together with features, and the receiver exploits its own LiDAR observation for feature fusion. The encoder uses a density-preserving deep point cloud compression method with “multi-embedding downsampling,” the channel encoder uses self-attention, and the feature fusion module uses cross-attention. The method also introduces “the nonlinear activation layer and transfer learning to improve the training of deep neural networks in the presence the digital channel noise” (Liu et al., 8 Sep 2025).

A distinctive property of this LPC-FT system is that it is designed for practical digital transmission rather than only analog DeepJSCC. The digital chain includes quantization, modulation, channel noise, demodulation, and a straight-through estimator. To stabilize this training, the method inserts a nonlinear activation such as $h(\cdot)$9 before the digital pipeline. Reported results show that LPC-FT is “more robust and effective than traditional octree-based compression followed by channel coding,” and “reducing the Chamfer Distance by 30% and improving the PSNR by 1.9 dB on average” relative to the compared methods (Liu et al., 8 Sep 2025).

A related but more explicitly task-driven formulation appears in “Point Cloud Feature Coding for Object Detection over an Error-Prone Cloud-Edge Collaborative System” (Tian et al., 4 Mar 2026). Here, the edge device extracts multi-scale detector features

$g(\cdot)$0

then transmits only a compacted version of the deepest feature. The source encoder performs “removing task-irrelevant regions and applying channel-wise dimensionality reduction to task-relevant areas,” the geometry part is protected by “LDPC + BPSK,” the attribute part is sent through an “SNR-adaptive channel encoder,” and the cloud reconstructs missing shallow-layer features by a “diffusion-based feature upsampling module” (Tian et al., 4 Mar 2026). On KITTI, the method reports “a 172-fold reduction in feature size” while maintaining “3D average precision scores of 93.17%, 86.96%, and 77.25% for easy, moderate, and hard objects, respectively, over a 0 dB SNR wireless channel” (Tian et al., 4 Mar 2026).

System	Transmitted representation	Channel strategy
SEPT	Global latent vector $g(\cdot)$1 from geometry features	DeepJSCC over AWGN (Bian et al., 2023)
Local-global SemCom	Lossy local features $g(\cdot)$2, lossless centroids $g(\cdot)$3 and global feature $g(\cdot)$4	JSCC for $g(\cdot)$5, strong coding for $g(\cdot)$6 (Xie et al., 2024)
SoM-aided LPC-FT	Density-preserving compressed point set with fused features	Self-attention channel encoding, cross-attention fusion (Liu et al., 8 Sep 2025)
Cloud-edge feature coding	Compacted detector feature point cloud	LDPC for geometry, SNR-adaptive JSCC for attributes (Tian et al., 4 Mar 2026)

Taken together, these works establish LPC-FT as both a geometry-reconstruction problem and a task-split inference problem. In the former, features are transmitted to recover a point set; in the latter, features are transmitted so that the cloud can continue the detector.

5. Relation to compression, densification, and temporal prediction

LPC-FT is embedded in a larger ecosystem of LiDAR communication methods. “Point Cloud in the Air” organizes this broader field into four frameworks: “Semantic communication framework based on DeepJSCC (SEPT),” “Representational compression framework inspired by NeRF,” “Uplink point cloud feature aggregation framework,” and “Distributed broadcast framework for delay-critical applications” (Shao et al., 2024). This classification is useful because it makes clear that feature transmission is only one design point among raw point-cloud compression, representational coding, over-the-air aggregation, and delay-oriented scheduling.

Raw-geometry compression remains a strong baseline. “PointCompress3D” targets roadside LiDARs and reports “a frame rate of 10 FPS while keeping compression sizes below 105 Kb, a reduction of 50 times, and maintaining object detection performance on par with the original data,” with “a PSNR d2 of 94.46 and a BPP of 6.54” on the TUMTraf dataset family (Zimmer et al., 2024). Likewise, “Multiscale Latent-Guided Entropy Model for LiDAR Point Cloud Compression” introduces hierarchical latent variables and reports “state-of-the-art performance” on SemanticKITTI and Ford, together with “more than 99.8% of decoding time” savings compared to previous state-of-the-art methods (Fan et al., 2022). More recent lossless and near-lossless predictive schemes such as “Re-Densification Meets Cross-Scale Propagation” and “LiZIP” emphasize compact feature representations for efficient predictive coding and V2X suitability, with the former achieving “26 FPS for both encoding and decoding at 12-bit quantization” and the latter reporting “a 7.5%-14.8% reduction in file size compared to the industry-standard LASzip” (Yu et al., 28 Aug 2025, Shibu et al., 24 Mar 2026).

Two adjacent lines of work serve as receiver- or sender-side companions to LPC-FT rather than direct feature codecs. “Real-Time LiDAR Point Cloud Densification for Low-Latency Spatial Data Transmission” reconstructs dense depth maps from sparse LiDAR and RGB by a joint bilateral filtering pipeline, yielding “full HD resolution in real time (30 fps)” and “~23 ms per frame” on RTX 4090, “over 15× faster than a recent training-based depth completion approach” (Murasaki et al., 3 Jan 2026). “RAI-Net: Range-Adaptive LiDAR Point Cloud Frame Interpolation Network” positions interpolation explicitly as a tool “for reducing the amounts of point cloud transmission” and “could be integrated into any LiDAR point cloud compression systems for inter prediction” (Zhao et al., 2021). This suggests that practical LPC-FT systems may place learned reconstruction either before transmission, after transmission, or between transmitted key states.

A recurrent boundary in this literature is therefore not between “communication” and “compression,” but between transmitting raw geometry, transmitting a geometric representation, and transmitting task-ready features. LPC-FT occupies the last two positions most strongly.

6. Evaluation criteria, misconceptions, and open issues

The dominant reconstruction metrics in LPC-FT are Chamfer distance and MPEG geometry PSNR. SEPT and the local-global semantic communication model both optimize Chamfer distance,

$g(\cdot)$7

while evaluating D1 and D2 PSNR (Bian et al., 2023, Xie et al., 2024). In the 2024 semantic communication model, D1 is derived from point-to-point MSE and D2 from point-to-plane error using the bounding-box diagonal $g(\cdot)$8 (Xie et al., 2024). For feature transmission aimed at detection, the relevant metric shifts to downstream AP, as in the cloud-edge feature coding work (Tian et al., 4 Mar 2026).

A common misconception is that geometric fidelity alone is sufficient. “On the Impact of LiDAR Point Cloud Compression on Remote Semantic Segmentation” explicitly rejects this view by proposing a semantic distortion metric

$g(\cdot)$9

where $$\mathbf{y}' = \mathbf{z}' + \mathbf{w}, \quad \mathbf{w}\sim \mathcal{CN}(0,N_0),$$0 is the set of nearest-neighbor pairs involving label $$\mathbf{y}' = \mathbf{z}' + \mathbf{w}, \quad \mathbf{w}\sim \mathcal{CN}(0,N_0),$$1, and $$\mathbf{y}' = \mathbf{z}' + \mathbf{w}, \quad \mathbf{w}\sim \mathcal{CN}(0,N_0),$$2 is the subset with a label mismatch (Fernandes et al., 27 Sep 2025). For the human class, the metric is weighted further: $$\mathbf{y}' = \mathbf{z}' + \mathbf{w}, \quad \mathbf{w}\sim \mathcal{CN}(0,N_0),$$3 Using this task-aware criterion, the paper concludes that maintaining high segmentation quality requires “approximately 0.6 MB/s for G-PCC and 2.8 MB/s for L3C2” in the tested remote segmentation setting (Fernandes et al., 27 Sep 2025). Within LPC-FT, this functions as a benchmark: any feature-transmission scheme that claims superiority over raw point-cloud codecs should be judged not only by D1, D2, or Chamfer distance, but also by task-conditioned distortion.

The main limitations reported across LPC-FT studies are consistent. Several systems are validated on synthetic CAD models rather than driving LiDAR; several use AWGN rather than vehicular fading; several optimize only reconstruction loss rather than downstream task objectives; and several do not explicitly exploit temporal redundancy (Bian et al., 2023, Xie et al., 2024). The cloud-edge feature coding work addresses task performance directly, but under AWGN and for a single detector family (Tian et al., 4 Mar 2026). This suggests that future LPC-FT research will likely move toward temporally aware, task-conditioned, multi-agent, and channel-realistic designs, while retaining the current central principle: transmit the smallest representation that preserves the geometry or semantics actually needed at the receiver.