- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes SoDPCC, a learned compression framework for dynamic LiDAR point clouds. The main idea is to jointly represent two consecutive frames using a shared octree constructed from their quantized union, while encoding frame membership using temporal states. The shared-octree occupancy is modeled by GeoContextNet, and the temporal states are compressed by TemporalNet using progressive grouped state factorization (PGSF) and state-geometry coupling (FSGC). The method is evaluated on SemanticKITTI and Ford, where it reports substantial BD-rate reductions over G-PCC and several learned compression baselines. Ablation studies are provided to analyze the contributions of the shared-octree representation and the proposed temporal-state modeling components.
Critical strengths
- Strong compression performance. SoDPCC demonstrates substantial compression gains over existing methods on both SemanticKITTI and Ford. In particular, the method achieves large BD-rate reductions relative to G-PCC across multiple distortion metrics and also consistently outperforms the compared learned compression methods. The ablation results further indicate that the shared-octree representation contributes substantially to the bitrate reduction by avoiding duplicated hierarchical geometry. Overall, the strong empirical compression performance is the main strength of this work and demonstrates the potential effectiveness of the proposed representation.
Critical weaknesses
-
Limited practicality due to high computational cost and inherent look-ahead latency. Despite being evaluated only at relatively limited octree depths (), SoDPCC requires several seconds per frame for both encoding and decoding (up to 5.33 s and 5.07 s at , respectively). This is particularly concerning for dynamic LiDAR applications, where low-latency processing is often important. In addition, the proposed non-overlapping two-frame joint coding requires the subsequent frame before the current pair can be encoded, introducing an inherent one-frame look-ahead delay. The combination of substantial computational latency and mandatory frame buffering significantly limits the practical applicability of the proposed framework.
-
Insufficient evaluation of the shared-octree representation across quantization resolutions. The main compression gain of SoDPCC relies on sharing hierarchical geometry between consecutive frames, and the paper itself shows that reduced inter-frame overlap enlarges the shared octree and degrades coding efficiency. However, the experiments are restricted to only three octree depths (). As the depth increases and voxelization becomes finer, exact voxel overlap between consecutive LiDAR scans may decrease due to sampling variation, ego-motion, and object motion, especially since no ego-pose registration or motion compensation is applied before constructing the shared octree. The paper does not analyze the shared-voxel ratio as a function of depth or demonstrate that the reported gains persist in higher-resolution regimes. This limitation is particularly important because it directly concerns the mechanism responsible for the main compression gain.
Questions for authors
-
Runtime and latency: Could the authors discuss the practical latency of SoDPCC in an online LiDAR setting? In particular, the reported encoding/decoding times already reach several seconds per frame at , and the two-frame joint representation additionally requires waiting for the subsequent frame. How does the overall end-to-end latency compare with the temporal baselines under the same look-ahead/buffering constraints, and how is the computational cost expected to scale at higher octree depths?
-
Sensitivity to quantization depth and inter-frame overlap: Since the effectiveness of the shared-octree representation depends on inter-frame voxel overlap, could the authors report the shared-voxel ratio (e.g., ) as a function of octree depth? More importantly, do the compression gains remain at finer quantization levels such as and ?
7 comments
Response to Q2
A2: Thank you for raising this point. We have extended the evaluation on SemanticKITTI and Ford to finer quantization levels, including D = 15 and 16, and report the shared-voxel ratio as a function of octree depth, together with the corresponding compression performance.
(1) Depth-dependent voxel overlap. We report the shared-voxel ratio (Voxel IoU) and hierarchical node-sharing ratio (Node Sharing), defined below. Here, and are the quantized voxel sets of two consecutive frames at depth , while , , and denote the non-leaf node counts of their independent and shared octrees, respectively.
As shown in Tables 3 and 4, exact voxel overlap decreases sharply with finer quantization, while hierarchical node sharing declines more gradually. For example, at D = 16, Voxel IoU falls to 0.002% on SemanticKITTI and 1.988% on Ford, whereas Node Sharing remains at 6.49% and 17.91%, respectively. Thus, shared-octree coding is not limited to merging coincident leaf voxels. Even when points in consecutive frames occupy different fine-resolution voxels, their octree paths can still share intermediate nodes and common ancestors, reducing repeated hierarchical geometry representation.
Table 3: Sensitivity to quantization depth on SemanticKITTI.
| Depth | Voxel IoU (%) | Node Sharing (%) | Geo. bpp ↓ | State bpp ↓ | Total bpp ↓ | D1 PSNR ↑ |
|---|---|---|---|---|---|---|
| 11 | 38.988 | 37.19 | 0.47 | 0.10 | 0.57 | 60.47 |
| 12 | 25.613 | 31.24 | 1.10 | 0.18 | 1.28 | 66.50 |
| 13 | 14.739 | 24.29 | 2.13 | 0.26 | 2.39 | 72.54 |
| 14 | 3.372 | 12.38 | 3.56 | 0.29 | 3.85 | 78.55 |
| 15 | 1.040 | 8.82 | 5.32 | 0.26 | 5.58 | 84.57 |
| 16 | 0.002 | 6.49 | 7.84 | 0.24 | 8.08 | 90.59 |
Table 4: Sensitivity to quantization depth on Ford.
| Depth | Voxel IoU (%) | Node Sharing (%) | Geo. bpp ↓ | State bpp ↓ | Total bpp ↓ | D1 PSNR ↑ |
|---|---|---|---|---|---|---|
| 11 | 47.342 | 41.03 | 1.47 | 0.34 | 1.81 | 58.54 |
| 12 | 28.980 | 36.92 | 3.03 | 0.46 | 3.49 | 64.33 |
| 13 | 13.476 | 31.47 | 4.31 | 0.47 | 4.78 | 70.45 |
| 14 | 6.329 | 25.76 | 5.87 | 0.48 | 6.35 | 76.23 |
| 15 | 3.296 | 21.26 | 7.65 | 0.41 | 8.06 | 82.36 |
| 16 | 1.988 | 17.91 | 10.01 | 0.45 | 10.46 | 88.06 |
(2) Rate-distortion performance at finer quantization depths. Tables 3 and 4 report compression performance across depths. At D = 15 and 16, SoDPCC achieves 5.58/8.08 bpp on SemanticKITTI, with D1 PSNR values of 84.57/90.59 dB. On Ford, the corresponding rates are 8.06/10.46 bpp, with D1 PSNR values of 82.36/88.06 dB. To further quantify compression gains, we have calculated D1 BD-rates relative to the G-PCC TMC13 v36.0 octree intra anchor in two quality regions: conventional quality, defined as 50–80 dB, and high quality, defined as 80 dB and above.
As shown in Table 5, SoDPCC achieves D1 BD-rate reductions of 41.94% and 45.77% on SemanticKITTI and Ford in the conventional-quality region, and retains reductions of 40.02% and 38.65% in the high-quality region, respectively. Although the gains decrease at higher quality, substantial compression benefits remain despite the reduced exact voxel overlap. Together with the node-sharing statistics, these results support the effectiveness of hierarchical geometry reuse and learned temporal-state coding at finer quantization levels.
Table 5: D1 BD-rate (%) relative to the G-PCC TMC13 v36.0 in conventional-quality and high-quality regions.
| Dataset | Conventional-quality BD-rate ↓ | High-quality BD-rate ↓ |
|---|---|---|
| SemanticKITTI | -41.94% | -40.02% |
| Ford | -45.77% | -38.65% |
Follow-up on Practical Codec Trade-offs
Thank you for the detailed rebuttal and additional experiments. The authors have addressed most of my original concerns, particularly regarding quantization-depth sensitivity and runtime analysis. I appreciate these efforts.
I understand that SoDPCC primarily targets offline storage, where longer encoding times may be acceptable. However, decoding cost remains relevant even in this setting, and the trade-off between compression gains and computational overhead is still worth considering.
While the compression benefits of the shared-octree representation are convincing, I would appreciate the authors' perspective on the practical implications of pairwise merged-frame coding, particularly regarding frame-level accessibility and decoding dependencies compared with conventional frame-wise codecs.
I am not requesting additional experiments, but simply a brief discussion of these design trade-offs. This clarification would help me better assess the remaining concerns before making my final decision on the rating.
Response to Q1
A1: Thank you for the question. We have supplemented the evaluation on SemanticKITTI with a complete encoding and decoding runtime breakdown, including preprocessing, neural inference, entropy coding, and end-to-end processing time. We have also extended the timing evaluation to higher octree depths.
(1) End-to-end processing time. We use rosbag to replay point cloud sequences at 10 Hz, simulating the sequential arrival of LiDAR frames. At this frequency, SoDPCC’s two-frame joint representation introduces approximately 0.1 s of additional look-ahead delay while waiting for the subsequent frame. Compared with the current processing times of several seconds, computation remains the dominant cost.
To compare computational costs under consistent conditions, all methods use the same input-buffering conditions. We separately measure preprocessing, neural inference, and entropy coding, together with the end-to-end processing time for encoding and decoding, as shown in Table 1. All methods are tested on an Intel Xeon Gold 6234 CPU (3.30 GHz) and an NVIDIA RTX 4090 GPU. SoDPCC’s times are averaged over jointly coded frame pairs.
Table 1. Runtime breakdown on SemanticKITTI at D = 14. Times are in seconds per input frame.
For the complete pipeline, SoDPCC achieves lower encoding/decoding times than STAEM (9.71/7.89 s versus 10.17/9.92 s), but remains slower than DAPCC (5.44/5.36 s). The additional cost relative to DAPCC mainly arises from four-stage temporal-state modeling and repeated full-spatial context propagation in TemporalNet. In return, SoDPCC achieves the lowest bitrate of 3.85 bpp, reducing the rate by 12.9% and 6.8% relative to STAEM and DAPCC, respectively. Our study focuses on exploiting inter-frame dependencies through shared geometry and learned temporal-state modeling for efficient offline storage. In this setting, where compression efficiency takes priority over processing speed, the bitrate savings offer a favorable trade-off for the additional computational cost.
(2) Runtime scaling with octree depth.
To further evaluate computational cost at different reconstruction qualities, we have extended the runtime evaluation to higher octree depths. Table 2 summarizes the end-to-end processing times of encoding and decoding, together with bitrate and rate-distortion performance.
As expected, the end-to-end encoding and decoding times increase with octree depth and spatial resolution. This increase is common in octree-based learned entropy models, as finer quantization generally produces longer occupancy sequences and increases the computational workload. For example, STAEM’s encoding/decoding times also increase several-fold, from 3.34/2.66 s at D = 12 to 24.88/23.53 s at D = 16. The higher computational cost accompanies improved reconstruction quality, with SoDPCC’s D1 PSNR increasing from 78.55 dB at D = 14 to 84.57/90.59 dB at D = 15/16. At D = 14, the quantization precision is already at the centimeter level, with the coordinate-wise error bounded by (eq<214−1400≈0.0244 m). Evaluation at such moderate octree depths is also common practice in related studies, including STAEM, DAPCC, OctFormer, DuOctree, and TopNet. The additional D = 15 and 16 experiments extend this range and explicitly evaluate the computational cost of finer reconstruction. We hope these results address your concern.
Table 2. End-to-end processing time and D1 PSNR at different octree depths on SemanticKITTI. Times are in seconds per input frame.