Unsupervised Learning of Shape and Pose with Differentiable Point Clouds

Eldar Insafutdinov and Alexey Dosovitskiy.
Neural Information Processing Systems, 2018.

Website: https://eldar.insafutdinov.com/PointClouds
Paper: https://arxiv.org/abs/1810.09381

Setup

Install dependencies

The code is in Python 3.6 and Tensorflow >= 1.11. We provide instructions on how to install all dependencies with conda, but using pip should also work. We suggest downloading miniconda for python 3.

Create Python 3.6 environment (some packages are not yet available for the more recent python 3.7):

conda create -n py36 python=3.6
conda activate py36

Install basic dependencies:

conda install numpy scikit-image pillow scipy scikit-learn pyyaml
pip install easydict

Install open3d (more information can be found on the official website):

conda install -c open3d-admin open3d

Install TensorFlow:

pip install tensorflow-gpu

Prepare training data

We use Shapenet v1 for all experiments in the paper.

For convenience, we provide the rendered images, used for training, for the chair, car and airplane classes of the ShapeNet dataset (category ID 03001627, 02958343 and 02691156 respectively):

cd data
./download_train_data.sh 03001627

Convert training images to TFRecords format:

./create_tf_records.sh 03001627

Download pre-computed ground truth point clouds for evaluation:

./download_ground_truth.sh 03001627

You can also generate ground truth yourself as described here.

Train and Evaluate

To train and evaluate the full model without camera pose supervision execute the following:

cd experiments/chair_unsupervised
# train and compute accuracy on the validation set
python ../../dpc/run/train_eval.py
# compute accuracy on the test set
python ../../dpc/run/predict_eval.py --eval_split=test

You can use a --gpu flag to specify an ID of the GPU you want to run on.

The file chamfer_pred_test.txt contains the accuracy of 3D shape reconstruction represented by the two quantities: coverage and precision. Chamfer distance metric is the sum of precision and coverage. More details on the evaluation metric can be found in the paper.

The file pose_error_pred_test.txt contains camera pose estimation error. The first quantity is accuracy at the 30° threshold and the second one is the median error in degrees.

A configuration file to train a model with camera pose supervision is located in experiments/chair_camera_supervision.

Visualise

We provide a Jupyter notebook to visualise predicted shapes. The rendering code uses Blender, which you can install in the external/ subdirectory under name blender or simply create a symlink, for example:

ln -s /path/to/blender-2.79b external/blender

After, you can lunch the notebook experiments/visualise.ipynb.

Citation

@inproceedings{insafutdinov18pointclouds,
title = {Unsupervised Learning of Shape and Pose with Differentiable Point Clouds},
author = {Insafutdinov, Eldar and Dosovitskiy, Alexey},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2018}
}