Hyperspherical Embedding for Point Cloud Completion

Junming Zhang · Haomeng Zhang · Ram Vasudevan · Matthew Johnson-Roberson

GENERATED ILLUSTRATION

TL;DR

Projecting encoder embeddings onto a unit hypersphere makes point cloud completion training more stable and improves results in single-task and multi-task learning.

BRIEF

Depth sensors usually capture only part of an object, so point cloud completion tries to predict the full shape from that partial view. Most methods use an encoder that compresses the input into an embedding, which a decoder turns into a shape. These embeddings tend to be sparsely spread in feature space, which hurts generalization. The authors add a module that normalizes embeddings onto a unit hypersphere, so only direction is optimized. They show more stable training across learning rates and more compact embeddings, with consistent gains in single-task and multi-task settings.

Abstract

Most real-world 3D measurements from depth sensors are incomplete, and to address this issue the point cloud completion task aims to predict the complete shapes of objects from partial observations. Previous works often adapt an encoder-decoder architecture, where the encoder is trained to extract embeddings that are used as inputs to generate predictions from the decoder. However, the learned embeddings have sparse distribution in the feature space, which leads to worse generalization results during testing. To address these problems, this paper proposes a hyperspherical module, which transforms and normalizes embeddings from the encoder to be on a unit hypersphere. With the proposed module, the magnitude and direction of the output hyperspherical embedding are decoupled and only the directional information is optimized. We theoretically analyze the hyperspherical embedding and show that it enables more stable training with a wider range of learning rates and more compact embedding distributions. Experiment results show consistent improvement of point cloud completion in both single-task and multi-task learning, which demonstrates the effectiveness of the proposed method.