acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

LiDAR point clouds provide explicit, deterministic physical boundaries critical for collaborative safety-critical perception. However, modern wireless channels are subject to dynamic bandwidth fluctuations and transient link outages. While most existing LiDAR codecs assume ideal channels and collapse under transmission disruptions, robust alternatives (such as deep JSCC or MDC) rely on statistical estimation, turning exact physical measurements into unverified algorithmic estimates. To address this dilemma, we propose Proteus, a learned LiDAR codec operating on 2D range images. Built upon a progressive bit-plane slicing representation, Proteus inherently ensures that stream truncation mathematically maps to deterministic spatial precision degradation without geometric hallucinations. To reconcile high-efficiency bits-back coding with stream-level truncation robustness, Proteus decouples the frame representation into two independent coding pathways: (i) the significant range bit-planes (SIG), which employs a self-contained bits-back coder equipped with an Autoregressive Initial Bits (ArIB) mechanism to eliminate single-frame overhead while securing an essential perceptual lower bound; and (ii) the insignificant range bit-planes and attributes (INS), which serializes the remaining low range planes before subordinate attributes via FIFO range coding, gracefully shedding attribute precision first during bandwidth drops. Extensive experiments on the Waymo Open Dataset and SemanticKITTI demonstrate that Proteus tolerates up to approximately 70% bitstream truncation, while outperforming established standards (e.g., G-PCC) and the representative learned compressor Unicorn under ideal channel conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LiDAR Point Cloud Compression, Bits-Back Coding, Truncation Robustness, Bit-Plane Slicing, Progressive Compression, Collaborative Perception

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wXzKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a LiDAR range image compression method called Proteus which is robust to bitstream truncation. The method dedouples the bitstream into two part (significant, and insignificant range bit-planes) and compresses independently. The method encodes the significant range bit-planes (SIG) with bits-back coder and the insignificant range bit-planes (INS) with FIFO range coding. In SIG, the mid sub-block is encoded and used as initial bits for coding of the head sub-block. In INS which can be truncated, the remaining bit-planes are ordered before the attribute bit-planes, so that the truncated range is zero-padded and the truncated attribute is predicted by a prediction head. Experiments on Waymo and SemanticKITTI show performance gains over previous G-PCC and Unicorn.

Critical strengths

The paper clearly formulates the problem and aligns with the goal of decoding the bitstream under truncated scenario.

The proposed method shows promising compression performance over previous methods.

Critical weaknesses
  1. Limited bitrate range. In real world scenario, the bandwidth for LiDAR transmission is often limited. But the SIG is fixed to the first 10 bit-planes, so the method cannot decode when the bitrate is lower than the size of SIG.

Q1. Please provide results with different split index (e.g., smaller SIG, such as the first 5), and compare with G-PCC and Unicorn at low bitrate such as 1 to 5 bpp.

  1. Unclear novelty. The proposed method is built on existing techniques. The bit-plane progressive coding corresponds to JPEG 2000, and bits-back coding corresponds to ArIB-BPS. The main difference from ArIB-BPS seems to be that the initial bits are moved from INS to the mid sub-block, which is hard to regrad as technical novelty.

Q2. Please clarify whether the first OP in Figure 5 is SIG coding only with INS truncated. If so, it shows the performance of the main contribution of the paper. Please compare the performance with ArIB-BPS, G-PCC, and Unicorn under the SIG only setting (without INS) by changing the split index (e.g., smaller SIG, larger SIG).

Questions for authors

See the Q1 and Q2.

Reviewer_mduvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Proteus, a LiDAR compression method that allows the bitstream to be truncated while still reconstructing the point cloud. The main idea is to separate essential geometry information (SIG) from additional information (INS), so the decoder can progressively improve reconstruction quality as more bits are received. The authors also use an attribute reconstruction module to improve intensity quality when only part of the bitstream is available.

The method is evaluated on SemanticKITTI and Waymo, with comparisons against existing codecs and ablation studies. The results show good compression performance and demonstrate that the proposed design supports progressive reconstruction under bitstream truncation

Critical strengths

The separation of SIG and INS allows the decoder to progressively reconstruct LiDAR data when the enhancement bitstream is truncated. This is a useful design for scalable transmission, although its robustness relies on the complete reception of SIG.

Critical weaknesses
  1. Limited novelty: the separation of SIG from INS is straightforward.
  2. Limited validation of robustness: The truncation experiments assume that SIG is always fully received, while only INS can be truncated. This demonstrates progressive decoding under controlled conditions, but does not fully support the paper's broader claims about robustness to packet loss and unstable network conditions.
  3. Insufficient comparison with LiDAR-specific methods: The evaluation mainly focuses on general-purpose compression codecs, without sufficient comparison against existing LiDAR-specific methods. This makes it difficult to assess whether Proteus offers a meaningful improvement over prior work designed for the same type of data and application
Questions for authors
  1. The truncation experiments assume that SIG is always fully received. Could the authors clarify how this assumption relates to realistic wireless transmission conditions, particularly when packet loss or bandwidth fluctuations affect the essential SIG data?

  2. While the authors explain the lack of directly comparable truncation-robust codecs, why are LiDAR-specific intra-frame compression methods not included in the rate–distortion comparisons under ideal channel conditions?

Reviewer_Tg2yOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Proteus, a learned LiDAR compression method based on range images and progressive bit-plane coding. It separates significant geometry bits from less significant geometry and attribute bits, enabling truncated bitstreams to produce valid reconstructions. The method also uses bits-back coding mechanism to improve compression efficiency.

Critical strengths
  1. The truncation-robust design is clearly motivated. By organizing geometry and attribute information into progressive bit planes, Proteus can decode truncated streams while preserving a valid reconstruction.

  2. The method combines robustness with an efficient entropy-coding design. The decoupled SIG/INS architecture and ArIB-based bits-back coding aim to retain compression efficiency while supporting progressive transmission.

Critical weaknesses
  1. Some results require further verification
  • In Fig. 5(a), for WOD geometry compression with the D1 metric, the JPEG XL curve has an unexpected S-shaped trend. Moreover, its final rate point appears to drop from roughly 110 dB to 100 dB, which looks anomalous. The authors should verify the data and the plotting procedure.

  • For WOD attribute compression, Draco performs substantially better than G-PCC. This is somewhat surprising because Draco is generally designed with fast encoding&decoding in mind, often with a trade-off in compression efficiency. In contrast, on SemanticKITTI, G-PCC performs better than Draco and JPEG XL. The appendix provides a possible explanation based on dataset statistics, but the discrepancy is large enough that more detailed analysis and reproducibility information are needed.

  1. The comparison is not sufficiently complete
  • The paper lacks comparison with methods that are closer in design to Proteus, especially learned range-image compression methods such as RangeCM. RangeCM is discussed in the latency analysis, but there is no corresponding rate-distortion comparison.

  • Fig. 4 only shows Proteus under progressive bitstream truncation. It would be useful to show reconstructions from other methods at comparable bitrates, so that the visual advantage of Proteus can be evaluated fairly.

  • In Fig. 5, Proteus appears to be evaluated mainly above 3 bpp. Its performance at lower bitrates is important for the claimed communication setting and should be reported.

  1. The experimental settings need to be further clarified
  • Table 1 does not clearly specify the precision used by every baseline. Some methods are point-cloud codecs, such as G-PCC, Draco, and Unicorn, while JPEG XL and Proteus operate on image or range-image representations. The authors should explain how the data were quantized and aligned for each method.

If the authors can adequately address these concerns, I would be happy to increase my score.

Questions for authors

Please see the weaknesses above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.