normal_estimation

Normal Estimation on PCPNet Dataset

This repo imiplements the normal estimation on PCPNet dataset presented in the paper. The experiment includes the comparison over four architectures:

1. PointNet: models/PointNet.py. Network architecture implementation is borrowed from here.
2. KPConv: models/KPConv.py. We implement the KPConv by ourselves according to the original KPConv paper.
3. DGCNN: models/DGCNN.py. Network architecture implementation is borrowed from here.
4. VecKM: models/VecKM.py. Our newly proposed $O(n)$ local geometry encoder.

Since computing the dense normal vectors from 100,000 points is not feasible for every architecture, we randomly select model.point_batch_size points out of the 100k points and compute their normals. Different architectures have different point_batch_size values. A higher point_batch_size value means larger scalability.

To reproduce the training, please download the PCPNet dataset from PCPNet and unzip them into the ./data folder. The file structure shall look like:

normal_estimation
├── data
│   └── PCPNet
│       ├── *.xyz
│       ├── *.normals
│       ├── list
│       └── npy
├── main.py
├── models
│   ├── DGCNN.py
│   ├── KPConv.py
│   ├── PointNet.py
│   └── VecKM.py
├── outputs
│   └── VecKM.out
└── README.md

Models

We get the following performance by training the models for three days. Further training can very likely further improve the performance. While this experiment is used for comparing the performance and inference speed, so we didn't train the models until their convergence, which may take far more time. You could check out the training logs at ./outputs/*.out.

	Average RMSE	Point Batch Size	Inference Time (100k points)
KPConv, #nbr=700, #kp=16	26.67	0.5k	5.942
DGCNN, #nbr=128	21.24	10k	OOM
PointNet, #nbr=700	20.27	1k	6.136
VecKM (Ours)	17.34	100k	0.149

The table is interpreted in this way: VecKM takes 0.149 second to compute the the normal vectors of 100k points, and similarly for others. So when the input size is 100k, VecKM is 40x faster than PointNet, and 41x faster than KPConv.

Training

python main.py --model VecKM
python main.py --model PointNet --num_neighbors 700
python main.py --model KPConv --num_neighbors 700 --kp 16
python main.py --model DGCNN --num_neighbors 128