[論文レビュー] Linked Dynamic Graph CNN: Learning on Point Cloud via Linking Hierarchical Features

Kuangen Zhang, Ming Hao+3 more|arXiv (Cornell University)|Apr 22, 2019

3D Shape Modeling and Analysis参考文献 24被引用数 142

ひとこと要約

LDGCNNは、複数の動的グラフから階層的特徴を結びつけ、変換ネットワークを排除することで、ModelNet40とShapeNetで最先端の結果を達成し、点群の分類とセマンティックセグメンテーションを改善します。

ABSTRACT

Learning on point cloud is eagerly in demand because the point cloud is a common type of geometric data and can aid robots to understand environments robustly. However, the point cloud is sparse, unstructured, and unordered, which cannot be recognized accurately by a traditional convolutional neural network (CNN) nor a recurrent neural network (RNN). Fortunately, a graph convolutional neural network (Graph CNN) can process sparse and unordered data. Hence, we propose a linked dynamic graph CNN (LDGCNN) to classify and segment point cloud directly in this paper. We remove the transformation network, link hierarchical features from dynamic graphs, freeze feature extractor, and retrain the classifier to increase the performance of LDGCNN. We explain our network using theoretical analysis and visualization. Through experiments, we show that the proposed LDGCNN achieves state-of-art performance on two standard datasets: ModelNet40 and ShapeNet.

研究の動機と目的

• Sparse, unordered point clouds without regular grids or normalsに直接学習を動機づける
• Improve upon DGCNN by linking features from different dynamic graphs
• Eliminate the transformation network and demonstrate MLP-based rotation invariance
• Enhance performance by freezing the feature extractor and retraining the classifier
• Provide theoretical analysis, visualization, and ablation to validate design choices

提案手法

• Construct a locally directed graph using K-NN on point clouds
• Extract local features with edge convolution using a shared MLP and max-pooling
• Link hierarchical features from different dynamic graphs to compute informative edge vectors
• Remove the transformation network and rely on MLPs to approximate rotation invariance
• Optionally freeze the feature extractor and retrain the classifier to boost performance

実験結果

リサーチクエスチョン

• RQ1 Can linking hierarchical features across dynamic graphs improve edge feature quality for point clouds?
• RQ2 Is a transformation network necessary for rotation/invariance when an MLP-based approach is used?
• RQ3 Does freezing the feature extractor and retraining the classifier improve overall accuracy on standard datasets?
• RQ4 How does LDGCNN compare to state-of-the-art methods on ModelNet40 classification and ShapeNet segmentation?
• RQ5 What is the impact of using K-NN in feature-space vs Euclidean space for edge computation?

主な発見

• LDGCNN achieves state-of-the-art performance on ModelNet40 for 1024-point input with OA 92.9% and MA 90.3%.
• On ShapeNet segmentation, LDGCNN attains competitive mean IoU across parts compared to prior methods (Table 2).
• Removing the transformation network while using MLPs for feature extraction maintains or improves performance.
• Freezing the feature extractor and retraining the classifier increases ModelNet40 OA from 91.8% to 92.9%.
• LDGCNN has a smaller model size and comparable or faster forward time than several baselines (e.g., OA 92.9% with 1.08M parameters).