PointNet For Geoerror Classification

This is a simplified model based on the PointNet and PointNet++ for geometry error classification.

Environment

1. PyTorch 1.5
2. Numpy
3. Matplotlib

Training

Specify the parameter MODEL in train.sh as pointnet2 for PointNet++ and pointnet for PointNet training.

sh train.sh

The training data is not uploaded. Follow the next command to test the model with fake random data instead.

Quick Test

# Check PointNet
python model/pointnet.py
# Check PointNet++
python model/pointnet_2.py

Guidance

First we need to initialize the training environment:

1. source env.sh

Then we start training with the specific training parameters in scripts/train.sh for each training

MODEL_TAG       # model tag for visualization
CLASSIFIER      # softmax or sigmoid
MODEL           # pointnet, pointnet2, bp
AC_FN           # sigmoid, relu
outf            # output dir

Start training with the following command:

sh scripts/train.sh

For visualizing the accuracy on both testing and training set on specific model, modify the following parameters in utils/ac_vis.py

accuracy_list
output_dir
# run visualization
python utils/ac_vis.py

For evaluation, specify the following parameters in the scripts/train.sh

PRETRAINED
MODE='testing'
BATCHSIZE=1