Hybrid-PCGC: Hybrid Point Cloud Geometry Compression Bridging Generalization-based Model and Implicit Neural Representation

Wenjie Huang; Qi Yang; Shuting Xia; He Huang; Zhu Li; Yiling Xu

cs.CVFeb 25, 2026

Hybrid-PCGC: Hybrid Point Cloud Geometry Compression Bridging Generalization-based Model and Implicit Neural Representation

Authors: Wenjie Huang, Qi Yang, Shuting Xia, He Huang, Zhu Li, Yiling Xu

Organizations: Shanghai Jiao Tong University · University of Missouri-Kansas City

Abstract

Efficient point cloud geometry compression (PCGC) is essential for 3D sensing applications such as autonomous driving and digital twins. Hybrid methods combine the generalization of learned models with the adaptability of Implicit Neural Representations (INRs), but existing schemes often use complex architectures, rigid update strategies, and post-training weight compression, limiting adaptation efficiency and increasing bitstream overhead. This paper proposes Hybrid-PCGC, which formulates hybrid compression as a joint geometry-and-model bitrate allocation problem and optimizes instance-specific weight offsets as rate-constrained transmitted variables. It separates dataset-agnostic prior modeling from instance-specific adaptation through four components: a Frozen General Module (FGM), a Dual Trainable Adaptive Module (DTAM), an Adaptive Scope Controller (ASC), and a Supervised Offset Compression Module (SOCM). FGM captures global geometric regularities, while the parameter-efficient DTAM adapts to individual point clouds. ASC selects the update scope based on coding difficulty, and SOCM jointly optimizes and compresses weight offsets through differentiable entropy supervision during training. Experiments show that Hybrid-PCGC balances fidelity and bitrate, reducing bitrate by 13.96% and 16.33% over G-PCC on lossless and lossy benchmarks, respectively, and saving up to 12.33% bitrate over the state-of-the-art hybrid method on challenging dense point clouds.

Figures & tables

Fig. 1 : Principle comparison of generalization-based, INR-based, and our Hybrid approaches, highlighting the model components utilized during the pretrain, instance adaptation, and weight coding phases.

Fig. 2 : Concept in octree.

Fig. 3 : (a) The framework includes the Pipeline. (b) The architecture of the FGM, which iteratively generates a robust prior. (c) The architecture of the DTAM, which consumes the prior from the FGM to refine the final occupancy prediction $P_{occ}^{i+1,j}$ .

Fig. 4 : Component of FGM. (a) FGM and ResNet. (b) Fusion module. It is used to integrate global information with the decoded local information. (c) Gate. It employs learnable weighting coefficients $M_{1}$ and $M_{2}$ to control the weighted sum of the two input features.

Fig. 5 : Principle of SOCM, where selected offset weights are quantized through STE and entropy-regularized by an entropy bottleneck. AE is an arithmetic encoder, AD is an arithmetic decoder.

Dataset	SparsePCGC	Unicorn	OctAttention	TopNet	GPCC	AnyPCC	AnyPCC-U	Ours	Ours-U
8iVFB	0.57	0.57	0.68	0.59	0.76	0.54	0.57	0.52	0.54
MVUB	0.69	0.69	0.76	0.69	0.94	0.67	0.75	0.65	0.70
Owlii	0.48	0.48	0.62	0.56	0.59	0.47	0.47	0.45	0.44
Thuman	1.70	1.70	2.31	2.20	2.00	1.58	1.64	1.56	1.59
ScanNet	1.86	1.86	2.13	2.03	2.03	1.83	1.88	1.78	1.81
KITTI	6.80	6.50	7.21	6.85	8.19	6.18	6.45	6.03	6.35

TABLE I: Lossless geometry compression performance on the benchmark. Bpp is reported for each dataset, with the best value in each row highlighted with a red background and the second-best value highlighted with a blue background. CR Gain is computed relative to G-PCC, and Enc/Dec denotes average encoding/decoding time. Ours and Ours-U correspond to the dataset-specific and unified settings, respectively. The results for Unicorn are cited directly from the original publication.

Fig. 6 : Performance comparison using rate-distortion curves. Comparisons on the human-body and ScanNet datasets exclude RENO, OctAttention, and TopNet, as they are specifically designed for lossy LiDAR compression and thus incompatible.

Dataset	UniPCGC	AnyPCC	AnyPCC-U	Ours	Ours-U
8iVFB	2.96%	-15.01%	-6.74%	-19.00%	-13.43%
MVUB	3.64%	-10.49%	0.04%	-16.14%	-4.29%
Owlii	1.50%	-20.45%	-10.72%	-19.26%	-13.96%
Thuman	-16.88%	-50.41%	-40.43%	-52.23%	-44.04%
ScanNet	/	-75.88%	-74.88%	-75.54%	-74.89%
Avg.	-2.20%	-34.45%	-26.54%	-36.44%	-30.12%

TABLE II : BD-Rate comparison on dense object and indoor point clouds using SparsePCGC as the anchor. Negative values indicate bitrate savings over SparsePCGC, and lower values are better.

Dataset	RENO	TopNet	AnyPCC	AnyPCC-U	Ours	Ours-U
KITTI	-12.31%	-16.81%	-23.37%	-20.07%	-26.83%	-23.19%
Ford	-8.84%	-17.55%	-18.60%	-17.47%	-21.72%	-20.64%
VGGT	9.75%	4.77%	-0.42%	-4.88%	-0.43%	-6.31%
Avg.	-3.80%	-9.86%	-14.13%	-14.14%	-16.33 %	-16.72%

TABLE III : BD-Rate comparison on LiDAR and reconstruction-derived point clouds using G-PCCv23 as the anchor. Negative values indicate bitrate savings over G-PCCv23, and lower values are better.

	Ours-U	Ours-U.M	CR-Gain
Ford	8.57	8.51	-0.70%
Dense	5.55	5.45	-1.80%
RS	3.50	3.4	-2.86%
NS	4.69	4.62	-1.49%
CS	3.18	2.96	-6.92%
Avg.	5.098	4.988	-2.16%

TABLE IV : Effect of using

M>1

point clouds in one adaptation unit. Ours-U denotes the unified model with

M=1

, while Ours-U.M shares one adapted offset among multiple point clouds.

Dataset	G-PCC	AnyPCC-U.B	Ours-U.B
8iVFB	0.57	0.56	0.54
Ford	9.77	8.57	8.55
GS	15.82	13.29	13.27
CS	4.94	3.36	3.28
Avg.	7.78	6.45	6.41
CR-Gain	0.00%	-17.11%	-17.56%

TABLE V : Comparison of unified base models without instance adaptation. AnyPCC-U.B and Ours-U.B denote the corresponding base models before optimizing instance-specific weight offsets.

Fig. 7 : Distribution evolution of transmitted weight offsets on CS during online instance adaptation. The blue histogram denotes the empirical frequency of quantized offsets, and the orange curve denotes the PMF estimated by SOCM.

Fig. 8 : Bitrate allocation across geometry scales and transmitted model offsets on representative datasets.

	IA	SOCM	ASC	Gain vs. Abla.3
Ours-U	✓	✓	✓	-7.21%
Abla.1	✓	✓	✗	-5.93%
Abla.2	✓	✗	✗	-5.66%
Abla.3	✗	✗	✗	0.00%

TABLE VI : Component ablation of the proposed unified model. IA denotes online instance adaptation. Gain is computed within this ablation study using Abla.3 as the reference.

Dataset	Ours-U	Abla.1	Abla.2	Abla.3
8iVFB	0.54	0.56	0.54	0.54
Ford	8.54	8.66	8.55	8.55
GS	11.7	11.7	12.14	13.27
CS	3.00	3.00	3.10	3.61
Gain vs. Abla.3	-7.21%	-5.93%	-5.66%	0.00%

TABLE VII : Detailed bitrate results for the ablation study. Gain is computed over the four representative datasets using Abla.3 as the reference.

Symbol	Description	Value
$C_{\mathrm{FGM}}$	Network channels used for FGM prior modeling in Stage I	64
$K_{\mathrm{FGM}}$	Convolution kernel size used for FGM prior modeling in Stage I	5
$C_{\mathrm{DTAM}}$	Network channels used for DTAM refinement in Stage II	8
$K_{\mathrm{DTAM}}$	Convolution kernel size used for DTAM refinement in Stage II	5
$B$	Batch size	1
$lr_{\mathrm{off}}$	Initial learning rate for offline training	0.0005

TABLE VIII: Training and online instance-adaptation parameters used in our experiments. Stage I, Stage II, and online instance adaptation correspond to the three-step training strategy described in the main paper. Dataset-dependent online settings follow the lossless testing protocol described in the main paper.

Dataset	Depth	Frames/Seq.	Total	Cond.	Avg. Points	Sequence/Source
8iVFB	10-bit	300	600	Easy	760375	Loot, Redandblack
MVUB	10-bit	/	461	Easy	1259792	Phil, Ricardo
Owlii	11-bit	600	2400	Easy	2635971	All sequences
Thuman	10-bit	/	20	Easy	441319	Frames 0500–0519
ScanNet	2 cm	100	100	Easy	126760	ScanNet test split
KITTI	1 mm	100	1100	Easy	119826	Sequences 11–21

TABLE IX : Test dataset protocol inherited from the AnyPCC benchmark. “Total” denotes the number of tested point clouds/frames. “/” means that the dataset is not organized by a fixed frames-per-sequence value in the benchmark table. Average point counts are reported before entropy coding under the benchmark preprocessing.

Checkpoint	Training data/source	Test use in Ours
8iVFB	8iVFB Longdress/Soldier and MVUB Andrew/David/Sarah training sequences	8iVFB Loot/Redandblack, MVUB Phil/Ricardo, and Owlii sequences
Thuman	Thuman training split, frames 0000–0499	Thuman test frames 0500–0519
ScanNet	ScanNet benchmark training scans with the same preprocessing	Disjoint ScanNet benchmark test split
KITTI	SemanticKITTI sequences 00–10, quantized from 1 mm coordinates	KITTI sequences 11–21 and the benchmark OOD sets mapped to the LiDAR-style checkpoint
Ford	Ford 01 training sequence with 1 mm quantization	Ford 02–03 test sequences
GS	Gaussian-splatting-derived training point clouds from the benchmark training source	Disjoint GS test set

TABLE X : Category-specific training and checkpoint selection used by Ours. Ours-U replaces all category-specific checkpoints in this table with one unified checkpoint.

Scope	Updated DTAM components	Offset params
$l=-1$	No instance adaptation; no DTAM offset is transmitted	0
$l=0$	Final linear layer of the occupancy-prediction MLP	528
$l=1$	MLPs involved in each DTAM stage	25,616
$l=2$	Stage-wise MLPs and their preceding SConv layers	31,008
$l=3$	Entire DTAM parameter set	44,952

TABLE XI: Definition of ASC adaptation scopes.

Field	Type	Length	Decoding dependency
Scope flag $l$	uint8	1 byte	Selects the DTAM offset tensor list and $N_{\Delta W_{l}}$
EB parameter count	uint32	4 bytes	Sanity check against fixed $N_{\mathrm{EB}}$
EB parameters	uint8 array	$N_{\mathrm{EB}}$ bytes	Dequantized by fixed per-tensor affine rules to compute the CDF
$\exp(F_{a})$	float16	2 bytes	Inverse scale for decoded offset symbols
Offset payload size	uint32	4 bytes	Number of bytes in the arithmetic-coded offset stream
Offset payload	byte array	$B_{\mathrm{AE}}$ bytes	Arithmetic decoded with the CDF computed from EB parameters

TABLE XII : Byte-level syntax of the model-offset sub-bitstream. The CDF is computed at the decoder from the transmitted entropy-bottleneck parameters and fixed codec constants; it is not transmitted as side information.

Point Cloud Compression

Point cloud compression reduces the storage and transmission cost of 3D points while preserving geometry, appearance, and task-relevant detail.

See all 22 papers

Towards Practical Lossless Neural Compression for LiDAR Point Clouds
Mar 26, 2026Pengpeng Yu, Haoran Li, Runqing Jiang +4Point Cloud Compression
Sun Yat-sen University, China · Pengcheng Laboratory, China
Geometry-Based Compression of Plenoptic Point Clouds
Aug 11, 2026Davi R. Freitas, Gustavo L. Sandri, Ricardo L. de QueirozPoint Cloud Compression
Inria - Rennes Bretagne-Atlantique Rennes, France · Instituto Federal de Bras´ılia Bras´ılia, Brazil · Universidade de Bras´ılia Bras´ılia, Brazil
Octree-based Learned Point Cloud Geometry Compression: A Lossy Perspective
Mar 30, 2026Kaiyu Zheng, Wei Gao, Huiming ZhengPoint Cloud Compression
School of Electronic and Computer Engineering, Shenzhen Graduate School, Peking University, Shenzhen, China

Mar 26, 2026J/K move · Enter open · S save

Towards Practical Lossless Neural Compression for LiDAR Point Clouds

Pengpeng Yu, Haoran Li, Runqing Jiang, Dingquan Li+3

Sun Yat-sen University, China · Pengcheng Laboratory, China

Point Cloud Compression

Open paper AskPDF arXiv

LiDAR point clouds are fundamental to various applications, yet the extreme sparsity of high-precision geometric details hinders efficient context modeling, thereby limiting the compression speed and performance of existing methods. To address this challenge, we propose a compact representation for efficient predictive lossless coding. Our framework comprises two lightweight modules. First, the Geometry Re-Densification Module iteratively densifies encoded sparse geometry, extracts features at a dense scale, and then sparsifies the features for predictive coding. This module avoids costly computation on highly sparse details while maintaining a lightweight prediction head. Second, the Cross-scale Feature Propagation Module leverages occupancy cues from multiple resolution levels to guide hierarchical feature propagation, enabling information sharing across scales and reducing redundant feature extraction. Additionally, we introduce an integer-only inference pipeline to enable bit-exact cross-platform consistency, which avoids the entropy-coding collapse observed in existing neural compression methods and further accelerates coding. Experiments demonstrate competitive compression performance at real-time speed. Code will be released upon acceptance. Code is available at https://github.com/pengpeng-yu/FastPCC.

Abstract

Point Cloud Compression

Explore similar work

Towards Practical Lossless Neural Compression for LiDAR Point Clouds

Geometry-Based Compression of Plenoptic Point Clouds

Octree-based Learned Point Cloud Geometry Compression: A Lossy Perspective