[Paper Review] A-CNN: Annularly Convolutional Neural Networks on Point Clouds

Artem Komarichev, Zichun Zhong+1 more|arXiv (Cornell University)|Apr 16, 2019

3D Shape Modeling and Analysis45 references37 citations

TL;DR

A-CNN introduces annular convolutions on point clouds using regular and dilated rings to capture local geometry, achieving state-of-the-art results on classification and segmentation benchmarks.

ABSTRACT

Analyzing the geometric and semantic properties of 3D point clouds through the deep networks is still challenging due to the irregularity and sparsity of samplings of their geometric structures. This paper presents a new method to define and compute convolution directly on 3D point clouds by the proposed annular convolution. This new convolution operator can better capture the local neighborhood geometry of each point by specifying the (regular and dilated) ring-shaped structures and directions in the computation. It can adapt to the geometric variability and scalability at the signal processing level. We apply it to the developed hierarchical neural networks for object classification, part segmentation, and semantic segmentation in large-scale scenes. The extensive experiments and comparisons demonstrate that our approach outperforms the state-of-the-art methods on a variety of standard benchmark datasets (e.g., ModelNet10, ModelNet40, ShapeNet-part, S3DIS, and ScanNet).

Motivation & Objective

• Motivate and address the challenge of learning from irregular and sparse 3D point clouds without converting to regular grids.
• Propose annular convolution to capture local neighborhood geometry via ring-shaped regions with regular and dilated rings.
• Develop a hierarchical network architecture that uses constraint-based ring k-NN, neighbor ordering, annular convolutions, and ring pooling.
• Demonstrate state-of-the-art performance on object classification, part segmentation, and semantic segmentation in large-scale scenes.

Proposed method

• Define regular and dilated rings to form non-overlapping local neighborhoods with constraint-based k-NN search.
• Order neighbors on a projected tangent plane to obtain a clockwise/counterclockwise sequence.
• Apply annular convolutions on the ordered ring neighborhoods with configurable kernel sizes (e.g., 1x3, 1x5) and cyclic concatenation for circular consistency.
• Pool features within each ring using max-pooling to create ring-wise descriptors, then concatenate rings for higher-level abstraction.
• Use a two-branch encoder with regular rings first and dilated rings second, followed by a classifier or segmentation head.
• Employ feature propagation via learned interpolation to upsample and predict per-point labels in segmentation tasks.

Experimental results

Research questions

• RQ1 Can annular convolutions on ring-neighborhoods improve the representation of local geometry in point clouds compared to existing point-based convolutions?
• RQ2 Do regular and dilated ring structures reduce overlap and redundancy while enabling multi-scale contextual aggregation in point clouds?
• RQ3 Does neighbor ordering on a projected tangent plane enable effective learning of local point relationships for 3D tasks?
• RQ4 Can the proposed A-CNN architecture achieve state-of-the-art results on classification, part segmentation, and semantic segmentation datasets?

Key findings

• A-CNN achieves competitive to state-of-the-art accuracy on ModelNet10 (AAC 95.3, OA 95.5) and ModelNet40 (AAC 90.3, OA 92.6) with 1024-point inputs.
• A-CNN outperforms many point-cloud based methods on model classification benchmarks and is competitive with, or surpasses, several volumetric and multi-view approaches.
• On ShapeNet-part, S3DIS, and ScanNet segmentation benchmarks, A-CNN attains strong mean IoU scores and competitive OA across datasets (as reported in Table 2 of the paper).
• Ablation studies show that rings without overlaps, neighbor ordering, and annular convolutions each contribute to performance, with the ring structure providing the largest gain.