AnyPcc: Compressing Any Point Cloud with a Single Universal Model

TL;DR

AnyPcc tackles the generalization gap in learning-based point cloud compression by introducing a Universal Context Model that fuses spatial and channel priors with a lightweight Instance-Adaptive Fine-Tuning strategy for per-instance adaptation. By unifying lossless and lossy compression, it achieves state-of-the-art performance across a broad 15-dataset benchmark, including strong OOD generalization with a single universal model (Ours-U). The approach leverages a coarse-to-fine, checkerboard-spatial and 8-bit occupancy code-based representation, and uses per-instance fine-tuning to optimize entropy models, balancing model cost against geometry bitrate savings. The results demonstrate robust, practical compression gains and reduced deployment overhead, with plans to release code and datasets to enable reproducible research.

Abstract

Generalization remains a critical challenge for deep learning-based point cloud geometry compression. We argue this stems from two key limitations: the lack of robust context models and the inefficient handling of out-of-distribution (OOD) data. To address both, we introduce AnyPcc, a universal point cloud compression framework. AnyPcc first employs a Universal Context Model that leverages priors from both spatial and channel-wise grouping to capture robust contextual dependencies. Second, our novel Instance-Adaptive Fine-Tuning (IAFT) strategy tackles OOD data by synergizing explicit and implicit compression paradigms. It fine-tunes a small subset of network weights for each instance and incorporates them into the bitstream, where the marginal bit cost of the weights is dwarfed by the resulting savings in geometry compression. Extensive experiments on a benchmark of 15 diverse datasets confirm that AnyPcc sets a new state-of-the-art in point cloud compression. Our code and datasets will be released to encourage reproducible research.

Figures (5)

• Figure 1: (a) AnyPcc Architecture. A single, unified model compresses point clouds from any source, with our Instance-Adaptive Fine-Tuning (IAFT) module boosting performance on out-of-distribution (OOD) data. (b) AnyPcc Benchmark. Our comprehensive benchmark features 15 diverse datasets, including both standard and extreme cases. When compared against five state-of-the-art methods, AnyPcc consistently achieves high compression efficiency across all types of point clouds.
• Figure 2: The superiority of our UCM in capturing contextual information across diverse point cloud types (e.g., dense and sparse).
• Figure 3: llustration of the proposed AnyPcc framework. AnyPcc formulates the compression task as a sequential prediction of multi-scale occupancy codes. It develops a Universal Context Model (UCM) that incorporates strong inductive biases from spatio-channel partitioning, enabling robust generalization across diverse data sources.
• Figure 4: Our Universal Context Model (UCM) uses a recursive, parameter-shared architecture. At each scale $l$, the model takes the voxels of that scale as input to predict their corresponding occupancy codes, repeating this process with a consistent structure across all scales, as illustrated in the transition from scale $l$ to $l+1$.
• Figure 5: 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.