acceptodds
Under review as a conference paper at ICLR 2027

SoDPCC: Shared-Octree Coding for Dynamic LiDAR Point Cloud Compression

Abstract

Dynamic LiDAR compression aims to compactly encode continuously evolving 3D scenes, while existing codecs mainly exploit temporal redundancy by compressing consecutive frames in a frame-wise manner. However, this paradigm leaves temporal correlation only implicitly reflected in the statistical similarity between adjacent encoded frames, which limits the effective exploitation of temporal information. To address this limitation, we propose SoDPCC, a learned LiDAR stream compression framework comprising a shared octree and inter-frame presence states, which respectively reduce spatial symbols through hierarchical geometry sharing and temporal uncertainty through compact state modeling. Specifically, SoDPCC constructs a shared octree for two consecutive frames to remove duplicated geometry across hierarchy levels, encodes each leaf’s inter-frame presence as a compact state, and models the resulting occupancy and state streams with dedicated Transformer-based entropy models. For temporal-state coding, we design Progressive Grouped State Factorization to progressively expose denser decoded state cues for subsequent prediction. We further propose Full-Spatial State-Geometry Coupling to integrate these cues with spatially aligned geometry features and propagate the resulting state-aware context over the complete leaf sequence. The experimental results show that our SoDPCC outperforms state-of-the-art methods, achieving 41.78% and 49.43% BD-Rate gains on SemanticKITTI and Ford datasets, respectively. Moreover, SoDPCC preserves competitive vehicle-detection performance on reconstructed LiDAR point clouds, demonstrating its practical utility for downstream perception.

Primary area:
applications to robotics, autonomy, planning
Keywords:
point cloud compression, autonomous driving, internet of vehicles

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6jh4OpenReview ↗
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
  1. 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
  1. Limited practicality due to high computational cost and inherent look-ahead latency. Despite being evaluated only at relatively limited octree depths (D=12,13,14D=12,13,14), SoDPCC requires several seconds per frame for both encoding and decoding (up to 5.33 s and 5.07 s at D=14D=14, 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.

  2. 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 (D=12,13,14D=12,13,14). 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
  1. 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 D=14D=14, 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?

  2. 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., ∣Vt∩Vt+1∣/∣Vt∪Vt+1∣\lvert V_t \cap V_{t+1}\rvert / \lvert V_t \cup V_{t+1}\rvert) as a function of octree depth? More importantly, do the compression gains remain at finer quantization levels such as D=15D=15 and D=16D=16?

7 comments
AuthorsOpenReview ↗

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.

MethodPreprocessing Enc./Dec.Inference Enc./Dec.Entropy Coding Overhead Enc./Dec.End-to-End Enc./Dec.bpp
STAEM2.07 / 1.555.69 / 5.822.41 /2.5510.17 / 9.924.42
DAPCC2.08 / 1.590.99 / 1.092.37 / 2.685.44 / 5.364.13
SoDPCC2.26/ 0.295.33 / 5.072.12 / 2.539.71 / 7.893.85

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<400214−1≈0.0244e_q < \frac{400}{2^{14}-1}\approx 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.

DepthEncoding time (s)Decoding time (s)bppD1 PSNR (dB)
123.652.761.2866.50
136.374.372.3972.54
149.717.893.8578.55
1518.5914.915.5884.57
1622.6315.708.0890.59
AuthorsOpenReview ↗

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, V0DV_0^D and V1DV_1^D are the quantized voxel sets of two consecutive frames at depth DD, while N0DN_0^D, N1DN_1^D, and NmDN_m^D denote the non-leaf node counts of their independent and shared octrees, respectively.

Voxel IoU(D)=∣V0D∩V1D∣∣V0D∪V1D∣,\mathrm{Voxel\ IoU}(D)=\frac{|V_0^D\cap V_1^D|}{|V_0^D\cup V_1^D|},

Node Sharing(D)=1−NmDN0D+N1D,\mathrm{Node\ Sharing}(D)=1-\frac{N_m^D}{N_0^D+N_1^D},

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.

DepthVoxel IoU (%)Node Sharing (%)Geo. bpp ↓State bpp ↓Total bpp ↓D1 PSNR ↑
1138.98837.190.470.100.5760.47
1225.61331.241.100.181.2866.50
1314.73924.292.130.262.3972.54
143.37212.383.560.293.8578.55
151.0408.825.320.265.5884.57
160.0026.497.840.248.0890.59

Table 4: Sensitivity to quantization depth on Ford.

DepthVoxel IoU (%)Node Sharing (%)Geo. bpp ↓State bpp ↓Total bpp ↓D1 PSNR ↑
1147.34241.031.470.341.8158.54
1228.98036.923.030.463.4964.33
1313.47631.474.310.474.7870.45
146.32925.765.870.486.3576.23
153.29621.267.650.418.0682.36
161.98817.9110.010.4510.4688.06

(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.

DatasetConventional-quality BD-rate ↓High-quality BD-rate ↓
SemanticKITTI-41.94%-40.02%
Ford-45.77%-38.65%
Reviewer_6jh4OpenReview ↗

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.

Reviewer_ALRmOpenReview ↗
Rating
2
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SoDPCC, which jointly encodes 2 consecutive LiDAR frames using a shared octree and temporal states. PGSF and FSGC are further introduced to improve temporal-state entropy modeling.

Critical strengths

The shared-octree representation is simple and effective in reducing geometric redundancy between 2 adjacent frames. The experiments are solid, with extensive baseline comparisons and ablation studies.

Critical weaknesses
  1. The improvements over MDLPCC are relatively limited, especially on Ford, while the encoding and decoding costs are considerably higher than DAPCC.
  2. The method only considers non-overlapping two-frame windows, leaving inter-window temporal correlations unexplored.
  3. The choice of 2-frame joint coding is not well justified, as no experiments with longer windows (3 or 4 frames) are provided.
Questions for authors

Could the authors provide results with longer coding windows (e.g., 3 or 4 frames)?

5 comments
AuthorsOpenReview ↗

Response to Q1 (Weakness 1)

A1: Thank you for pointing this out.

(1) Rate-distortion gains over MDLPCC and Ford performance. SoDPCC improves the D1/D2/CD BD-rates over MDLPCC by 2.35/2.62/4.96 percentage points on SemanticKITTI and 0.79/0.80/2.05 percentage points on Ford. To examine this dataset difference, we measured inter-frame voxel overlap across quantization depths. As shown in Table 1 below, Ford exhibits higher overlap at most evaluated depths. This greater spatial consistency allows temporal codecs to exploit inter-frame redundancy effectively, which may partly explain the smaller additional gains over MDLPCC on Ford.

Beyond these numerical gains, SoDPCC jointly codes shared geometry and explicit temporal states, providing a different coding strategy from frame-wise temporal prediction.

(2) Runtime cost and rate-distortion trade-off. At D = 14 on SemanticKITTI, SoDPCC requires 5.33/5.07 s per input frame for encoding/decoding inference, making it slightly faster than STAEM (5.69/5.82 s) but slower than DAPCC (0.99/1.09 s). The additional cost relative to DAPCC mainly arises from the four-stage temporal-state modeling and repeated full-spatial context propagation in TemporalNet.

In return, SoDPCC improves the D1/D2/CD BD-rates over DAPCC by 2.54/2.47/2.62 percentage points on SemanticKITTI and 6.98/7.22/9.34 percentage points on Ford. All improvements are differences between BD-rates measured against the same GPCC anchor. SoDPCC therefore trades additional computation for better compression performance, consistent with its intended application in offline storage.

Table 1: Inter-frame voxel overlap on SemanticKITTI and Ford.

DepthSemanticKITTI Voxel IoU (%)Ford Voxel IoU (%)
1138.98847.342
1225.61328.980
1314.73913.476
143.3726.329
151.0403.296
160.0021.988
AuthorsOpenReview ↗

Response to Q2 (Weakness 2&3)

A2: Thank you for raising this point. Existing dynamic LiDAR codecs, including STAEM, DAPCC, and MDLPCC, commonly use a two-frame prediction context, with the preceding reconstructed frame supporting current-frame occupancy prediction. The common goal is to exploit local temporal dependencies for effective entropy modeling. SoDPCC pursues this goal through a shared spatial representation, encoding frame membership as leaf-level temporal states and using PGSF and FSGC to integrate decoded states with spatial geometry features.

The current framework independently codes non-overlapping windows. Because their temporal states are defined on different shared voxel sets, cross-window modeling requires an additional mechanism for transferring context between these spatial supports. In this revision, we examine broader temporal dependencies by expanding the joint coding window, allowing more frames to be modeled within a single shared representation. Specifically, we have added experiments with 3-, 4-, and 5-frame joint coding on SemanticKITTI and compared them with the two-frame configuration in terms of representation overhead and rate-distortion performance, as detailed below.

(1) Representation overhead across coding-window lengths. To quantify the trade-off between geometry sharing and temporal-state overhead, we define an equivalent node count that expresses both costs in units of 8-bit octree occupancy symbols: Neq=Nmerged+Bstate8,N_{\mathrm{eq}}=N_{\mathrm{merged}}+\frac{B_{\mathrm{state}}}{8},

where NmergedN_{\mathrm{merged}} denotes the number of shared-octree occupancy nodes and BstateB_{\mathrm{state}} denotes the temporal-state representation size in bits, both averaged per input frame.

As shown in Table 2, longer windows do not necessarily yield a more compact representation. Although they enable greater hierarchical geometry sharing, the equivalent node count increases with window length at finer depths. At D = 14, it rises from 181,188.95 for two-frame coding to 198,751.84 for five-frame coding, an increase of 9.69%. These results show that the additional temporal-state overhead can outweigh the incremental geometry-sharing savings.

Table 2. Equivalent node counts per input frame for different coding-window lengths on SemanticKITTI.

DepthIndependent2-frame3-frame4-frame5-frame
91654.021468.461327.861291.951304.04
104765.084627.594417.244464.944646.43
1113741.6814193.8814128.3814654.2015558.23
1237625.4839777.4840506.1942463.9345371.84
1389496.3393764.8595761.3199836.50105843.66
14175693.88181188.95184150.59189927.82198751.84
15288474.35293462.72296564.04302795.30312990.89

(2) Rate-distortion performance across coding-window lengths. Table 3 compares the D1 and D2 BD-rates across coding-window lengths. Two-frame coding achieves the best performance, with D1/D2 BD-rates of -41.94%/-41.24%, while the gains progressively decrease for longer windows.

This trend may partly reflect the increasing complexity of temporal-state modeling. For a K-frame window, the number of possible nonempty frame-membership states grows as 2K−12^K-1, corresponding to 3, 7, 15, and 31 states for K = 2, 3, 4, and 5, respectively. The expanded state space captures more diverse temporal occupancy patterns, increasing the demands on conditional probability estimation. Together with the representation-overhead results in Table 2, these findings support two-frame coding as the most effective configuration among those evaluated, balancing geometry sharing, temporal-state overhead, and modeling complexity.

We have added this ablation study on coding-window length and the corresponding analysis to Section 4.3 of the revised manuscript.

Table 3: BD-rate (%) comparison of different coding-window lengths on SemanticKITTI, relative to GPCC.

Coding WindowD1 BD-rate ↓D2 BD-rate ↓
2-frame-41.94%-41.24%
3-frame-36.76%-35.57%
4-frame-30.57%-29.24%
5-frame-22.06%-20.55%
AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Thank you again for taking the time to review our work. As the discussion period draws to a close, we would be grateful to hear whether our response has addressed your concerns. We’d be happy to answer any further questions.

Reviewer_X4w7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes SoDPCC for dynamic LiDAR point cloud geometry compression. It jointly encodes two consecutive frames using a shared octree and frame-membership states, with Progressive Grouped State Factorization (PGSF) and Full-Spatial State-Geometry Coupling (FSGC) modeling state probabilities and geometric context. The paper reports compression gains over G-PCC and learning-based baselines and evaluates coding efficiency and downstream task performance.

Critical strengths
  1. The motivation for reducing redundant coding across adjacent frames through a shared structure is clear. The reported compression gains suggest practical value.
  2. The experiments cover traditional and learning-based baselines, coding efficiency, and downstream tasks, helping assess the method’s practical utility.
Critical weaknesses
  1. The core motivation has limited novelty. “Temporal Octrees for Compressing Dynamic Point Cloud Streams” already combines octrees from multiple frames into a shared structure, records frame membership using temporal bitmasks, and supports frame-group coding. The reported compression gains support the method’s effectiveness, but its contribution appears to lie more in LiDAR-specific design and learned probability modeling. The authors are encouraged to articulate these technical contributions more clearly.
  2. Experimental comparability requires clarification. SoDPCC jointly encodes two frames, but the GOP configurations and reference structures of the inter-frame baselines are insufficiently specified. This makes it difficult to assess how differences in temporal coding structures affect the comparison. Besides, reproducibility remains a concern. Encoder-decoder inconsistencies are found, and results on the Ford dataset after correction still differ from those reported. Missing pretrained weights and differences in training settings limit this comparison.
Questions for authors
  1. What are the substantive differences between the proposed shared representation and existing Temporal Octrees? Does the main additional contribution lie in the representation or in learned probability modeling?
  2. What GOP configurations, reference structures, and refresh intervals are used by the inter-frame baselines? Do the reported rates include all necessary overhead, and do efficiency comparisons distinguish processing time from the delay incurred by waiting for subsequent frames?
  3. Could the authors provide the weights and complete training and evaluation configurations used for the main results?
  4. The authors are encouraged to include comparisons with a more recent G-PCC version, such as TMC13 v34, using predictive-tree geometry coding.
6 comments
AuthorsOpenReview ↗

Response to Q1

We thank the reviewer for recognizing the clear motivation and practical value of our shared-octree framework, as well as the comprehensive evaluations covering compression efficiency and downstream tasks.

A1: Thank you for your questions. Our contribution lies not only in the organization of shared geometry and temporal states, but also in the development of learned entropy models specifically designed for this representation. SoDPCC represents consecutive frames using a shared octree with compact temporal states by directly merging quantized voxel sets before octree construction. This reduces duplicated hierarchical geometry coding and the number of temporal-state elements, rather than adopting the multi-level bitmask representation used in Temporal Octrees, which merges frame-wise octrees and attaches bitmasks across multiple hierarchical levels, introducing redundant representations and additional parent-child mask dependencies. SoDPCC also employs learned conditional probability modeling to capture temporal-state dependencies and their correlations with spatial geometry features, enabling more effective temporal-state entropy coding through PGSF and FSGC. The detailed differences are presented below.

(1) Shared representation and temporal-state organization. Temporal Octrees merges individual frame-wise octrees and attaches temporal bitmasks to nodes throughout the hierarchy, introducing hierarchical dependencies between parent and child masks. In contrast, SoDPCC directly merges quantized voxel sets and assigns frame-membership states only to the resulting leaf voxels before constructing the octree. This design substantially reduces the number of state-bearing elements, avoids maintaining hierarchical temporal-mask dependencies, and simplifies representation construction. Moreover, the unified voxel representation is not inherently restricted to octrees and can also be organized using sparse tensors.

(2) Learned temporal-state probability modeling. Temporal Octrees primarily exploits deterministic parent-child mask constraints to omit redundant temporal bits during binary serialization. In contrast, SoDPCC explicitly models the conditional probability distribution of temporal states given the reconstructed shared geometry. Specifically, PGSF progressively exploits dependencies among decoded states, while FSGC integrates these states with spatially aligned geometry features to improve probability estimation. The temporal-state coding ablations support the effectiveness of these designs. PGSF reduces the state bitrate from 0.66 to 0.29 bpp, while FSGC further reduces it from 0.45 to 0.29 bpp compared with SSCA under the same grouping strategy.

Thus, SoDPCC differs substantively in how temporal information is represented and modeled, rather than merely improving probability estimation within the existing Temporal Octrees representation. We have revised the Introduction and Related Work to distinguish the shared-geometry concept from our temporal-state organization and dedicated learned entropy modeling.

AuthorsOpenReview ↗

Response to Q2 and Q3

A2: Thank you for your question. We detail the temporal coding structures, timing definitions, and bitrate accounting below.

(1) GOP configurations and reference structures. SoDPCC adopts non-overlapping two-frame coding groups (GOP=2), where each pair is jointly encoded and independently decoded without inter-group references. In comparison, the evaluated dynamic LiDAR compression methods, including STAEM, DAPCC, and MDLPCC, do not employ conventional fixed-length GOPs with periodic key frames or representative-frame-based residual coding. Instead, the first frame of each sequence is independently encoded in intra/static mode, and each subsequent frame uses the immediately preceding fully reconstructed frame as its reference. These methods therefore follow a causal I0P1P2P3…I_0P_1P_2P_3\ldots structure without periodic intra refresh. We also evaluate GPCC TMC13 v36 in inter-frame predictive-tree mode on Ford, employing angular geometry prediction, forward inter-frame prediction, and global motion compensation, with bidirectional prediction disabled and randomAccessPeriod=32. We have specified the temporal coding configuration of each method, including its reference structure and refresh interval, in the revised manuscript.

(2) Bitrate accounting. All reported bitrates are calculated from the complete coded streams, including independently coded initialization frames, bitstream headers, and any transmitted side information required for decoding. The total coded bits are divided by the total number of input points in the evaluated sequence. For SoDPCC, this includes both the shared-geometry and temporal-state streams, as well as any independently coded unpaired final frame.

(3) Coding delay and processing time. Table 3 in the revised manuscript reports encoding and decoding inference time per input frame, rather than end-to-end latency. The reported times exclude the waiting time required to acquire subsequent frames, consistent with common practice in related studies, such as STAEM (ICME'23) and DAPCC (TGRS'25). SoDPCC primarily targets high-efficiency offline compression for storage rather than latency-sensitive real-time transmission. In this setting, consecutive frames are available before encoding. The primary objective of this study is to capture informative inter-frame dependencies and integrate them with spatial geometry features for effective entropy modeling, a challenging task for sparse and dynamically changing LiDAR sequences.

We have clarified the timing definition and application setting in the revised manuscript.

A3: Thank you for highlighting these reproducibility concerns. The initial supplementary materials focused on methodological details and experimental protocols but did not include all the resources needed for independent reproduction.

We have now provided the complete encoding and decoding implementations, pretrained weights, and the training and evaluation configurations used for the main results in an anonymous repository. The repository also includes instructions for reproducing the reported results on SemanticKITTI and Ford. Anonymous repository: https://anonymous.4open.science/r/SoDPCC_framework-7ABC/.

To investigate the reported discrepancy on Ford, could you please specify the encoder-decoder inconsistencies you observed and the corrections applied during reproduction? The relevant code locations, evaluation settings, and results would help us identify the source of the discrepancy.

AuthorsOpenReview ↗

Response to Q4

A4: Thank you for this suggestion. We have added comparisons using GPCC TMC13 v36.0, a more recent version than v34. The updated BD-rate results and rate-distortion curves are presented in Table 1 and Figure 4 of the revised manuscript, respectively.

Relative to the GPCC octree intra anchor, SoDPCC achieves D1/D2/CD BD-rate reductions of 41.94%/41.24%/41.50% on SemanticKITTI and 45.77%/44.67%/47.99% on Ford. On Ford, the predictive-tree intra and inter configurations achieve reductions of 18.52%/28.01%/32.85% and 18.94%/28.26%/33.66%, respectively, relative to the same anchor. These results show that SoDPCC retains its rate-distortion advantage when compared with the newer GPCC version, including its predictive-tree configurations. The specific coding configurations used for each dataset are detailed below.

(1) SemanticKITTI. We evaluate GPCC-Octree (intra) using conventional octree geometry coding. The five codingScale values are {0.05, 0.02, 0.01, 0.005, 0.002}, corresponding to spatial resolutions of {0.02, 0.05, 0.10, 0.20, 0.50} m, respectively. We set inputScale=1000 and disable inter-frame prediction and in-loop position quantization. The adopted GPCC common test conditions (CTC) do not provide predictive-tree intra or inter configurations for SemanticKITTI. Therefore, these two modes are evaluated only on Ford.

(2) Ford. We evaluate three configurations: GPCC-Octree (intra), GPCC-Pred (intra), and GPCC-Pred (inter). The octree configuration uses codingScale values of {0.001953125, 0.00390625, 0.015625, 0.03125, 0.125}, corresponding to spatial resolutions of {512, 256, 64, 32, 8} mm, respectively. Both predictive-tree configurations use angular geometry prediction with Ford’s 64-beam LiDAR calibration. We set positionBaseQp=58, with positionRadiusInvScaleLog2 values of {8, 7, 5, 4, 2} and corresponding positionAzimuthScaleLog2 values of {7, 7, 8, 9, 11}. The inter configuration additionally enables forward inter-frame prediction and global motion compensation, with corresponding interAzimScaleLog2 values of {6, 6, 7, 8, 10}. Bidirectional prediction is disabled, and randomAccessPeriod=32.

All configurations encode XYZ geometry only, and bitrate is measured in bits per input point. The baseline configurations are described in Section 4.1 of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.