Proteus: A Truncation-Robust Entropy Model for Progressive LiDAR Compression

Technical Summary: PROTEUS – A Truncation-Robust Entropy Model for Progressive LiDAR Compression

1. Problem Statement

LiDAR point clouds provide explicit, deterministic physical boundaries essential for safety-critical collaborative perception (e.g., collision detection, hazard avoidance). However, transmitting these dense data structures over dynamic wireless channels (V2X, UAV swarms) introduces two critical challenges:

• Bandwidth Fluctuation & Truncation: Transient link outages or bandwidth drops cause in-transit stream truncation. Conventional codecs collapse under such conditions, requiring latency-heavy feedback loops to renegotiate bitrates, which stalls real-time pipelines.
• Channel Impairments vs. Physical Fidelity: Existing robust frameworks, such as Deep Joint Source-Channel Coding (Deep JSCC) or Multiple Description Coding (MDC), handle packet loss by reconstructing missing regions via statistical or parametric estimation. While this maintains competitive perceptual quality or downstream mAP, it converts exact physical measurements into unverified algorithmic estimates. For safety-critical tasks, preserving the mathematically bounded, deterministic nature of the sensor's measurement—even at a coarser resolution—is indispensable.

The core challenge is designing a learned codec that achieves progressive, truncation-robust compression while strictly preserving the exact physical boundaries of the LiDAR measurements, avoiding the "hallucination" of geometry inherent in statistical estimation methods.

2. Methodology

The authors propose Proteus, a learned LiDAR codec operating on 2D range images. The framework decouples the frame representation into two independent coding streams to achieve stream-level truncation robustness:

A. Representation & Decoupling
The input range image $R = (x, a)$ consists of the range channel $x$ (bit depth $d$) and attribute channels $a$ (e.g., reflectance). The range channel is sliced into $d$ bit-planes and partitioned into:

1. SIG (Significant): The most significant bit-planes ($x_{1:s}$), comprising a "head" ($x_{1:h}$) and "mid" ($x_{h+1:s}$) sub-block.
2. INS (Insignificant): The remaining tail bit-planes ($x_{s+1:d}$) and all attribute channels.

B. SIG Coder: Self-Contained Bits-Back
The SIG block encodes the significant planes using a bits-back coding scheme that is entirely self-contained within the block.

• Mechanism: It utilizes an Autoregressive Initial Bits (ArIB) mechanism. The "mid" sub-block is encoded first without latent conditioning to generate initial bits. These bits are then used to decode the latent variable $z$ for the "head" sub-block, which is encoded under the posterior $q(z|x_{1:h})$.
• Benefit: This eliminates the single-frame overhead typically associated with bits-back coding (which usually requires amortization over a sequence). The SIG block is non-truncatable; it establishes a necessary, self-contained perceptual lower bound. If the stream is truncated before SIG is complete, the frame is undecodable, but if SIG is received, the fundamental geometric structure is preserved.

C. INS Coder: Truncation-Robust FIFO
The INS block encodes the insignificant range planes and attributes using a First-In-First-Out (FIFO) range coder.

• Ordering: The stream is serialized with range tail planes ($x_{s+1:d}$) preceding attribute planes ($a$). This prioritizes geometric precision over attribute fidelity during bandwidth drops.
• Truncation Properties:
  • Deterministic Precision Degradation: Truncating $N$ least significant bit-planes of the range is mathematically equivalent to zero-padding, resulting in a deterministic spatial precision degradation ($\Delta_N = 2^N \cdot \Delta_0$). This avoids the need for statistical estimation of missing geometry.
  • Hybrid Attribute Reconstruction: When attribute bits are truncated, the remaining bits are reconstructed via a hybrid lossless-predictive method. The decoded geometry (from SIG and the received INS range) serves as a strong structural prior for a prediction network $h_\theta$ to approximate the missing attribute details.
• Robustness: Since SIG accounts for ~30% of the bitstream, Proteus remains fully decodable even if up to 70% of the INS stream is truncated.

3. Key Contributions

1. Truncation-Robust Bits-Back: The paper addresses the incompatibility between truncation robustness and high-efficiency bits-back coding. By decoupling the stream and utilizing ArIB, Proteus enables progressive, graceful degradation with only a minor penalty in lossless compression efficiency, eliminating single-frame overhead.
2. Strategic Geometry-Attribute Ordering: A joint coding framework is designed where subordinate attribute data is placed after geometry in the truncatable stream. This allows the system to naturally sacrifice attribute precision first during bandwidth dips, maintaining downstream perception performance even with significant truncation.
3. Deterministic Physical Preservation: Unlike Deep JSCC or MDC, which rely on parametric estimation, Proteus ensures that truncated reconstructions are mathematically bounded projections of the source. The geometry is never "hallucinated"; it is simply reconstructed at a lower precision.

4. Experimental Results

Experiments were conducted on the Waymo Open Dataset (WOD) and SemanticKITTI.

• Truncation Robustness: Proteus tolerates up to ~70% bitstream truncation. Under truncation, geometric error remains mathematically bounded (e.g., truncating 3 planes yields ~8mm precision on KITTI), and downstream 3D object detection performance degrades progressively to a self-contained lower bound rather than collapsing.
• Compression Efficiency (Ideal Channels): Under ideal conditions (0% truncation), Proteus outperforms established standards (G-PCC, Draco, JPEG XL) and the learned baseline Unicorn.
  • SemanticKITTI: Achieves -35.41% BD-rate gain over G-PCC for geometry and -15.59% for intensity.
  • WOD: Achieves -42.72% BD-rate gain over G-PCC for geometry and -80.47% for intensity.
• Ablation Studies:
  • Removing the bits-back mechanism increases the rate by +20.19%.
  • Removing the ArIB scheme increases the rate by +31.17%.
  • The decoupled design incurs only a 6.56% efficiency tax compared to a coupled ArIB scheme to enable truncation robustness.
  • The hybrid attribute reconstruction module (LUT + CNN Head) is computationally lightweight (1.3 MB storage, 5.2 ms overhead) and critical for performance (removing both increases BD-rate by +67.8%).

5. Significance and Claims

The paper claims that Proteus resolves the fundamental tension between high-efficiency compression and channel-robust transmission for safety-critical LiDAR applications. By strictly restricting coding dependencies to intra-frame and decoupling the representation, Proteus achieves:

• Frame-level packet-loss resistance (no inter-frame prediction to break).
• Prefix-decodable bitstreams that yield valid reconstructions without prior coordination.
• Mathematically verifiable safety nets: The reconstructed point cloud retains its deterministic physical boundaries, ensuring that even under severe truncation, the data remains a faithful, albeit lower-precision, measurement rather than an unverified estimate.

The authors position Proteus as a solution that allows collaborative autonomous systems to share dense point clouds over unreliable wireless channels without compromising the mathematical integrity required for safety-critical decision-making.