With the rapid advancement of Augmented Reality (AR) and autonomous driving, the storage and transmission of high-density point clouds have become critical challenges. To address this, we propose HMD-PCGC, a novel end-to-end Point Cloud Geometry Compression framework that synergizes a Mamba-Diffusion architecture with an adaptive hybrid representation learning mechanism. Unlike traditional approaches that rely on single-modality inputs, our encoder integrates three complementary feature types—point-based, voxel-based, and implicit features—through a Cross-Representation Attention module. This design allows for the comprehensive capture of local geometric details and complex topologies. To effectively model long-range dependencies within the latent space without the computational bottleneck of Transformers, we employ a Mamba block (State Space Model) for sequential processing, coupled with a scorer-guided adaptive quantization module that dynamically allocates bit-rate based on feature significance. For the decoding stage, we utilize a conditional Diffusion Model that generates high-fidelity point clouds by reversing a noising process, conditioned on the compressed latent representation. Extensive experiments on the MPEG 8iVFB and MVUB datasets demonstrate that HMD-PCGC achieves significant improvements in compression efficiency. Specifically, our method yields average BD-Rate gains exceeding 90% against G-PCC (octree), 80% against G-PCC (trisoup), and 40% against V-PCC, while also outperforming several recent learning-based frameworks in both rate-distortion efficiency and visual reconstruction quality.

Despite the efficiency of lossy geometry-based point cloud compression, decompressed point clouds exhibit distortions that depend on the compression level. Severe distortions, in particular, highlight the need for geometry restoration. Existing methods are limited in practice: they train separate models for specific distortion levels and require prior knowledge of these levels at inference, preventing generalization to broader distortion cases. To address these issues, we propose GGRNet, a generalizable geometry restoration network for compressed point clouds. GGRNet trains a single model and requires no prior knowledge of distortion levels for inference. Specifically, it adopts a sparse encoder-decoder for hierarchical geometry reconstruction. To handle diverse point densities and distortions, we introduce distortion-aware mixture-of-experts descriptors, which adaptively capture global and local features via dynamic gating. We further equip the decoder with distortion prompt adapters to implicitly encode and adapt to different distortion levels. Finally, we design a perceptual alignment loss to ensure perceptual invariance in the learned prompts. Extensive experiments show that GGRNet achieves strong performance in both the geometry restoration and downstream analysis tasks, demonstrating practical potential for real-world applications.

In this paper, we mainly research lossy octree-based point cloud geometry compression. We analyze data characteristics of different point clouds and propose lossy approaches specifically (Fig. 1 (d-f)). For object point clouds that suffer from quantization step adjustment, we propose a new leaf nodes lossy compression method (Fig. 1 (a-b)), which achieves lossy compression by performing bit-wise coding and binary prediction on leaf nodes. For LiDAR point clouds, we discover the occupancy distribution similarity for octrees in the same depth. Therefore, we present variable rate approaches and propose a simple but effective rate control method. Experimental results demonstrate that the proposed leaf nodes lossy compression method significantly outperforms the previous octree-based method on object point clouds, and the proposed rate control method achieves about 1% bit error without finetuning on LiDAR point clouds.

Focusing on encoding the Region of Interest (ROI) in point clouds and allocating more bitstream is a crucial area of research. In processing point cloud data, the foreground ROI region typically contains critical information, making it essential for applications like autonomous driving and robot navigation. However, previous point cloud compression methods often treat the entire point cloud uniformly and fail to fully harness the significance of the ROI. This study is dedicated to preserving vital information in point cloud by optimizing the Point Cloud Compression (PCC) process and allocating more bitstream to the foreground ROI region. To achieve this goal, we introduce a Semantic-Aware Visual Decomposition Point Cloud Geometry Compression (SAVD-PCGC) strategy. It involves the initial identification of foreground and background regions, followed by allocating of additional bitstream resources to machine vision critical areas by controlling compression model parameters. We also propose corresponding compensation methods to reduce distortion loss in compression. This separation of foreground and background coding strategy aims to maintain compression performance while ensuring high-quality of the ROI region, thereby improving the performance of downstream tasks. Experimental results demonstrate that our approach significantly enhances the performance of point cloud object detection compared to traditional PCC methods.

Integrating compression with analysis for point clouds poses a formidable challenge due to the inherent tension between the primary goals of compression for a compact representation and analysis for rich semantic retention. To alleviate this gap and maximize the practical requirements, we introduce a Parallel Dual-branch Network (PDNet) for lossy point cloud geometry compression, whose outputs are also analysis-friendly. The proposed method uses a novel Transformer-based encoder-decoder framework to incorporate local and global attention for point cloud latent representation computation. Specifically, the encoder comprises a Multi-scale Local-Global Feature extraction (MLGF) block to capture compact local and global latent features. The decoding and the hyper-prior modules employ a Transformer with No Position Embedding (TNPE) block and a Multilayer Perceptron (MLP) layer to reconstruct point clouds accurately. Furthermore, our method allows simultaneous point cloud analysis based on the compressed bitstream, such as point cloud classification. Experimental results demonstrate that our PDNet achieves nearly a 40% BD-Rate gain compared to G-PCC and other point-based compression counterparts. Besides, a 26% accuracy improvement in instance classification is observed compared to reconstructed point cloud classification.

Learning-based point cloud compression technology based on voxel structure has made significant progress in recent years. However, these methods have to train multiple models for different bit rates, which consume more storage and training resources. In a limited bit rate environment, these methods can only choose one model from a finite number of pretrained models, resulting in bit waste and low compression efficiency. To tackle these difficulties, we propose a variable-rate point cloud geometry compression network. A channel feature adjustment module (CFA) is designed to regulate features and achieve variable bit rates within one model. Then, in order to compress the point cloud to any bit rate, we also propose feature and entropy parameters interpolation methods. The proposed framework can realize a fine rate interval of 0.0001 bits per point (BPP) in point cloud compression. The experiment results demonstrate that the variable-rate network shows comparable performance to the other voxel based fixed-rate point cloud compression methods.

For exploring the Rate-Distortion-Complexity (RDC) optimization in point cloud compression, we propose a point cloud compressor with dynamic channel. In the transform process of the proposed compressor ( Fig 1.a ), we devise a sparse convolution operator, named AdaSConv, shown in Fig 1.b , to support RDC optimization, which ensures model capacity can adjust Rate-Distortion performance. What is more, to fill the blank of improved entropy model in point cloud feature compression, we design a 3D checkerboard entropy model. The 3D checkerboard divides points in the whole space into two parts: anchor and non-anchor, which will be compressed in sequence. As Fig 1.c illustrates, the compression of non-anchor will refer to the information in coded anchor points through Masked AdaSConv ( Fig 1.d ). We conduct floating point operations (FLOPs) computation, which reveals that the smallest rate point only consumes 7% of the FLOPs used by the full-width model. Besides, experiment results show 3D checkerboard has at most 13.86% gains of BD-Rate compared with factorized entropy model in the same experimental settings in Owill dataset with only slight extra time and computation.

Facing the explosion of massive amount of high precision depth maps, we pro-pose a novel end-to-end lossless compression method for high precision depth maps. The whole process is comprised of two sub-processes, named pre-processing of depth maps and deep lossless compression of processed depth maps. The deep lossless compression network consists of two sub-networks, named lossy compression net-work and lossless compression network. We leverage the concept of pseudo-residual to guide the generation of distribution for residual and avoid introducing context models. Our end-to-end lossless compression network outperforms tested learned and non-learned codecs with at least 7% reduction on real world datasets (DIODE and SementicKITTI) and has low computational cost. The visualization of the whole operations is shown as Figure 1. More details and code are available at https://git.openi.org.cn/OpenCompression/DeepLosslessCompression.