Here we present code to build an autoencoder for point clouds, with PointNet encoder and various kinds of decoders. We train and test our autoencoder on the ShapeNetPart dataset. This is a side project I played with recently -- you are welcomed to modify it for your own projects or research. Let me know if you discover something interesting!
This repository is under the MIT license. See the LICENSE file for detail.
We need TensorFlow (version>=1.4).
For point cloud reconstruction loss function, we need to compile two custum TF operators under tf_ops/nn_distance (Chamfer's distance) and tf_ops/approxmatch (earth mover's distance). Check the tf_compile_*.sh script under these two folders, modify the TensorFlow and CUDA path accordingly before you run the shell script to compile the operators. Check this PAPER for an introduction for these two point cloud losses.
For a visualization helper, go to utils/ and run sh compile_render_balls_so.sh -- run python show3d_balls.py to test if you have successfully compiled it.
ShapeNetPart dataset is available HERE (635MB). Simply download the zip file and move the shapenetcore_partanno_segmentation_benchmark_v0 folder to data.
To visualize the dataset, run (type q to go to the next shape, see show3d_balls.py for more detailed hot keys):
python part_dataset.py
To train the most basic autoencoder (fully connected layer decoder with Chamfer's distance loss) on chair models with aligned poses, simply run the following command:
python train.py --model model --log_dir log_chair_norotation --num_point 2048 --category Chair --no_rotation
You can check more options for training by:
python train.py -h
To test and visualize results of the trained autoencoder above, simply run:
python test.py --model model --model_path log_chair_norotation/model.ckpt --category Chair
You can check more options for testing by:
python test.py -h