xnuohz / KPConv-dgl

A DGL implementation of "KPConv: Flexible and Deformable Convolution for Point Clouds" (ICCV 2019).

DGL Implementation of KPConv

This DGL example implements the GNN model proposed in the paper KPConv: Flexible and Deformable Convolution for Point Clouds. For the original implementation, see here.

Contributor: xnuohz

Requirements

The codebase is implemented in Python 3.7. For version requirement of packages, see below.

dgl 0.6.0.post1
torch 1.7.0
logzero 1.7.0

The dataset used in this example

ModelNet10 for classification. Dataset summary:

• Number of point clouds: 3,991(train), 908(test)
• Number of classes: 10

ModelNet40 for classification. Dataset summary:

• Number of point clouds: 9,843(train), 2,468(test)
• Number of classes: 40

Usage

Note: we only support KPConv rigid in this example.

Train a model which follows the original hyperparameters

# ModelNet10
python main.py --epochs 100

# ModelNet40
python main.py --data-type large --epochs 60

Performance

Dataset	ModelNet10	ModelNet40
Result(Paper)	-	92.9
Result(Author)	92.60	91.6
Result(DGL)	92.18	86.1

Speed

Dataset	ModelNet10	ModelNet40
Result(Author)	46.38	59.87
Result(DGL)	304.20	908.24

Issue

• 991 point clouds in modelnet40 train set will cause dgl._ffi.base.DGLError: Expect number of features to match number of nodes (len(u))
• means nodes and positions do not match in kpconv nn graph or pool bipartite graph
• error idxs are saved into error_idx.npy