Robust Point Cloud Processing through Positional Embedding
This work addresses robustness issues in point cloud processing for applications like detection and alignment, though it appears incremental as it builds on existing embedding methods.
The paper tackles the sensitivity of learned per-point embeddings in 3D point cloud processing to out-of-distribution noise and outliers by introducing an analytical per-point embedding based on bandwidth, connecting it to positional embedding like random Fourier features, and demonstrates robust results in tasks such as classification and registration with various noise categories.
End-to-end trained per-point embeddings are an essential ingredient of any state-of-the-art 3D point cloud processing such as detection or alignment. Methods like PointNet, or the more recent point cloud transformer -- and its variants -- all employ learned per-point embeddings. Despite impressive performance, such approaches are sensitive to out-of-distribution (OOD) noise and outliers. In this paper, we explore the role of an analytical per-point embedding based on the criterion of bandwidth. The concept of bandwidth enables us to draw connections with an alternate per-point embedding -- positional embedding, particularly random Fourier features. We present compelling robust results across downstream tasks such as point cloud classification and registration with several categories of OOD noise.