Dynamic Local Feature Aggregation for Learning on Point Clouds

Zihao Li,Pan Gao,Hui Yuan,Ran Wei

DOI: https://doi.org/10.48550/arXiv.2301.02836

2023-01-07

Computer Vision and Pattern Recognition

Abstract:Existing point cloud learning methods aggregate features from neighbouring points relying on constructing graph in the spatial domain, which results in feature update for each point based on spatially-fixed neighbours throughout layers. In this paper, we propose a dynamic feature aggregation (DFA) method that can transfer information by constructing local graphs in the feature domain without spatial constraints. By finding k-nearest neighbors in the feature domain, we perform relative position encoding and semantic feature encoding to explore latent position and feature similarity information, respectively, so that rich local features can be learned. At the same time, we also learn low-dimensional global features from the original point cloud for enhancing feature representation. Between DFA layers, we dynamically update the constructed local graph structure, so that we can learn richer information, which greatly improves adaptability and efficiency. We demonstrate the superiority of our method by conducting extensive experiments on point cloud classification and segmentation tasks. Implementation code is available: https://github.com/jiamang/DFA.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper aims to address the problem of learning from point cloud data, particularly improving performance in classification and segmentation tasks. Specifically, existing point cloud learning methods aggregate features of neighboring points by constructing graphs in the spatial domain, leading to feature updates at each layer based on fixed spatial neighbors. To overcome this limitation, the paper proposes a Dynamic Feature Aggregation (DFA) method.

#### Main Contributions:
1. **Dynamic Feature Aggregation (DFA) Module**: Constructs local graph structures by finding the k-nearest neighbors in the feature domain and dynamically updates these graph structures to improve adaptability and efficiency.
2. **Relative Position Encoding and Semantic Feature Encoding**: Utilizes relative position information and semantic feature information for encoding to explore potential positional and feature similarity information.
3. **Efficient Feature Extraction**: Combines multi-level and multi-source features to effectively represent objects for point cloud classification and segmentation tasks.

#### Method Overview:
- **DFA Layer**: Finds the k-nearest neighbors in the feature domain and extracts rich local features through relative position encoding and semantic feature encoding.
- **Network Architecture**: Adopts a structure similar to PointNet, extracting low-dimensional global features and fusing them with local features to generate high-dimensional global features for classification and segmentation tasks.
- **Dynamic Graph Update**: Reconstructs local graph structures in the feature domain each time to explore more potential positional information.

Through these methods, the paper conducts extensive experiments on multiple benchmark datasets, demonstrating its superiority.