[Paper Review] Deep AutoEncoder-based Lossy Geometry Compression for Point Clouds

Wei Yan, Yiting Shao+4 more|arXiv (Cornell University)|Apr 18, 2019

3D Shape Modeling and Analysis29 references59 citations

TL;DR

It proposes a deep autoencoder architecture for lossy geometry compression of point clouds that directly takes point clouds as input and achieves superior rate–distortion performance vs. MPEG TMC13 (average 73.15% BD-rate gain).

ABSTRACT

Point cloud is a fundamental 3D representation which is widely used in real world applications such as autonomous driving. As a newly-developed media format which is characterized by complexity and irregularity, point cloud creates a need for compression algorithms which are more flexible than existing codecs. Recently, autoencoders(AEs) have shown their effectiveness in many visual analysis tasks as well as image compression, which inspires us to employ it in point cloud compression. In this paper, we propose a general autoencoder-based architecture for lossy geometry point cloud compression. To the best of our knowledge, it is the first autoencoder-based geometry compression codec that directly takes point clouds as input rather than voxel grids or collections of images. Compared with handcrafted codecs, this approach adapts much more quickly to previously unseen media contents and media formats, meanwhile achieving competitive performance. Our architecture consists of a pointnet-based encoder, a uniform quantizer, an entropy estimation block and a nonlinear synthesis transformation module. In lossy geometry compression of point cloud, results show that the proposed method outperforms the test model for categories 1 and 3 (TMC13) published by MPEG-3DG group on the 125th meeting, and on average a 73.15\% BD-rate gain is achieved.

Motivation & Objective

• Motivate and enable flexible, data-driven geometry compression for irregular 3D point clouds.
• Develop an end-to-end autoencoder framework that directly processes raw point clouds without voxelization or image-like representations.
• Incorporate a learnable entropy model and a differentiable quantization scheme to optimize rate–distortion.
• Demonstrate competitive compression performance across common object categories against a MPEG PCC baseline.

Proposed method

• Use a pointnet-based encoder to obtain a permutation-invariant latent representation.
• Apply a uniform quantizer (with straight-through or uniform-noise approximation during training) to produce a discrete latent code.
• Incorporate an entropy estimation (entropy bottleneck) module to model the prior of latent codes for rate estimation.
• Optimize a rate–distortion objective combining Chamfer distance distortion with estimated bit-rate.
• Use a fully-connected decoder to reconstruct the 3D point cloud from the latent code.
• Train and evaluate on ShapeNet to compare against MPEG TMC13 across multiple object categories.

Experimental results

Research questions

• RQ1 Can an end-to-end autoencoder learn efficient lossy geometry representations directly from point clouds without voxelization?
• RQ2 What gain in rate–distortion performance can be achieved over MPEG PCC’s TMC13 when using a learned entropy model and differentiable quantization?
• RQ3 How does incorporating an entropy bottleneck influence bitrate and reconstruction quality for point cloud geometry?
• RQ4 Is the approach robust across different object categories (e.g., chair, airplane, table, car) under varying bitrates?

Key findings

• The proposed autoencoder-based geometry codec outperforms TMC13 across chair, airplane, table, and car categories at all tested bitrates.
• On average, the method achieves a 73.15% BD-rate gain over TMC13.
• An ablation study shows the entropy bottleneck module provides a 19.3% BD-rate gain over a baseline without entropy estimation.
• The reconstructed point clouds tend to be denser and achieve lower bits-per-point at similar PSNR than TMC13.