Learning Convolutional Transforms for Lossy Point Cloud Geometry Compression (1903.08548v2)

Abstract: Efficient point cloud compression is fundamental to enable the deployment of virtual and mixed reality applications, since the number of points to code can range in the order of millions. In this paper, we present a novel data-driven geometry compression method for static point clouds based on learned convolutional transforms and uniform quantization. We perform joint optimization of both rate and distortion using a trade-off parameter. In addition, we cast the decoding process as a binary classification of the point cloud occupancy map. Our method outperforms the MPEG reference solution in terms of rate-distortion on the Microsoft Voxelized Upper Bodies dataset with 51.5% BDBR savings on average. Moreover, while octree-based methods face exponential diminution of the number of points at low bitrates, our method still produces high resolution outputs even at low bitrates. Code and supplementary material are available at https://github.com/mauriceqch/pcc_geo_cnn .

• The paper proposes a novel data-driven lossy point cloud compression method using a 3D convolutional auto-encoder and binary classification for decoding.
• Experimental results demonstrate that the learned convolutional transforms achieve an average 51.5% BDBR savings over the MPEG anchor on the MVUB dataset.
• This CNN-based approach maintains higher resolution geometry at low bitrates compared to traditional methods, promising improved efficiency for VR and MR applications.

Learning Convolutional Transforms for Lossy Point Cloud Geometry Compression

The paper "Learning Convolutional Transforms for Lossy Point Cloud Geometry Compression" introduces a data-driven method for the lossy compression of static point cloud geometries, utilizing convolutional neural networks (CNNs) as the key mechanism for compression. This research focuses on tackling the challenges presented by the vast size and complexity of point clouds which are crucial for applications in Virtual Reality (VR) and Mixed Reality (MR).

Methodology and Approach

The authors have proposed a novel compression approach rooted in the learning of convolutional transforms. The main components of the method include:

1. Convolutional Auto-Encoder: The compression technique hinges on a 3D convolutional auto-encoder, which comprises analysis and synthesis transforms. These transforms are trained to create compact representations of the original point cloud data. The network is designed to process 3D voxels directly, eliminating the reliance on predefined transformation techniques like wavelets.
2. Uniform Quantization: Quantization is accomplished through a method that adds noise during training to approximate the effects of discrete entropy coding. This aspect ensures that the network's outputs remain differentiable and thus trainable.
3. Binary Classification for Decoding: The decoding of compressed data is treated as a binary classification problem, determining the occupancy state of each voxel within a grid. This novel perspective enables effective geometric reconstructions even at lower bitrates.
4. Rate-Distortion Optimization: The compression framework incorporates a trade-off parameter that jointly optimizes for both rate and distortion, enabling better control over the quality and size of the compressed data.

Experimental Results

The paper reports on rigorous experimentation using established datasets such as ModelNet40 for training, and the Microsoft Voxelized Upper Bodies (MVUB) dataset for testing. The proposed method achieves significant compression efficiency:

• Rate-Distortion Performance: The method demonstrates an average 51.5% BDBR savings compared to the MPEG reference anchor on the MVUB dataset. This underlines the method's superiority in maintaining high precision in geometric reconstruction at any given bitrate.
• Resolution Retention at Low Bitrates: A critical advantage of the proposed CNN-based method is its ability to maintain higher resolution outputs even at lower bitrates, unlike traditional octree-based approaches which tend to lose geometric details significantly as bitrate decreases.

Implications and Future Directions

This research introduces a new potential direction in point cloud compression by leveraging the adaptability and learning capabilities of neural networks. The implications are considerable, promising more usable VR and MR experiences given the improved efficiency in storing and transmitting point cloud data.

For future work, the approach could be expanded to include dynamic point clouds and attribute-based data compression, as well as potentially integrating further into an end-to-end learnable pipeline for 3D data processing tasks. Moreover, optimization of loss functions and quantization thresholds in the neural network framework could lead to even finer control over the balance between rate and distortion, catering to specific application requirements.

Overall, the paper presents a comprehensive framework that underscores the capabilities of modern machine learning techniques in addressing longstanding challenges in point cloud compression, offering a substantial contribution to the field of computational geometry and multimedia processing.