Skip to content


Switch branches/tags


Failed to load latest commit information.
Latest commit message
Commit time

PointCNN: Convolution On X-Transformed Points

Created by Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen.


PointCNN is a simple and general framework for feature learning from point cloud, which refreshed five benchmark records in point cloud processing (as of Jan. 23, 2018), including:

  • classification accuracy on ModelNet40 (91.7%, with 1024 input points only)
  • classification accuracy on ScanNet (77.9%)
  • segmentation part averaged IoU on ShapeNet Parts (86.13%)
  • segmentation mean IoU on S3DIS (65.39%)
  • per voxel labelling accuracy on ScanNet (85.1%)

See our preprint on arXiv (accepted to NeurIPS 2018) for more details.

Pretrained models can be downloaded from here.

Performance on Recent Benchmarks

Revisiting Point Cloud Classification: A New Benchmark Dataset and Classification Model on Real-World Data

PartNet: A Large-scale Benchmark for Fine-grained and Hierarchical Part-level 3D Object Understanding

ABC: A Big CAD Model Dataset For Geometric Deep Learning

Practical Applications

3D cities: Deep Learning in three-dimensional space (from Esri)

PointCNN: replacing 50,000 man hours with AI (from Esri)

Point Cloud Segmentation using PointCNN in ArcGIS API for Python (from Esri)

More Implementations

We highly welcome issues, rather than emails, for PointCNN related questions.


Our code is released under MIT License (see LICENSE file for details).

Code Organization

The core X-Conv and PointCNN architecture are defined in

The network/training/data augmentation hyper parameters for classification tasks are defined in pointcnn_cls, for segmentation tasks are defined in pointcnn_seg.

Explanation of X-Conv and X-DeConv Parameters

Take the xconv_params and xdconv_params from for example:

xconv_param_name = ('K', 'D', 'P', 'C', 'links')
xconv_params = [dict(zip(xconv_param_name, xconv_param)) for xconv_param in
                [(8, 1, -1, 32 * x, []),
                 (12, 2, 768, 32 * x, []),
                 (16, 2, 384, 64 * x, []),
                 (16, 6, 128, 128 * x, [])]]

xdconv_param_name = ('K', 'D', 'pts_layer_idx', 'qrs_layer_idx')
xdconv_params = [dict(zip(xdconv_param_name, xdconv_param)) for xdconv_param in
                 [(16, 6, 3, 2),
                  (12, 6, 2, 1),
                  (8, 6, 1, 0),
                  (8, 4, 0, 0)]]

Each element in xconv_params is a tuple of (K, D, P, C, links), where K is the neighborhood size, D is the dilation rate, P is the representative point number in the output (-1 means all input points are output representative points), and C is the output channel number. The links are used for adding DenseNet style links, e.g., [-1, -2] will tell the current layer to receive inputs from the previous two layers. Each element specifies the parameters of one X-Conv layer, and they are stacked to create a deep network.

Each element in xdconv_params is a tuple of (K, D, pts_layer_idx, qrs_layer_idx), where K and D have the same meaning as that in xconv_params, pts_layer_idx specifies the output of which X-Conv layer (from the xconv_params) will be the input of this X-DeConv layer, and qrs_layer_idx specifies the output of which X-Conv layer (from the xconv_params) will be forwarded and fused with the output of this X-DeConv layer. The P and C parameters of this X-DeConv layer is also determined by qrs_layer_idx. Similarly, each element specifies the parameters of one X-DeConv layer, and they are stacked to create a deep network.

PointCNN Usage

PointCNN is implemented and tested with Tensorflow 1.6 in python3 scripts. Tensorflow before 1.5 version is not recommended, because of API. It has dependencies on some python packages such as transforms3d, h5py, plyfile, and maybe more if it complains. Install these packages before the use of PointCNN.

If you can only use Tensorflow 1.5 because of OS factor(UBUNTU 14.04),please modify "isnan()" to "std::nan()" in "/usr/local/lib/python3.5/dist-packages/tensorflow/include/tensorflow/core/framework/numeric_types.h" line 49

Here we list the commands for training/evaluating PointCNN on classification and segmentation tasks on multiple datasets.

  • Classification

    • ModelNet40

    cd data_conversions
    python3 ./ -d modelnet
    cd ../pointcnn_cls
    ./ -g 0 -x modelnet_x3_l4
    • ScanNet

    Please refer to for downloading ScanNet task data and scannet_labelmap, and refer to for downloading ScanNet benchmark files:


    |_ data

    |_ scannet_labelmap

    |_ benchmark

    cd ../data/scannet/scannet_dataset_download/
    mv ./scannet_labelmap/scannet-labels.combined.tsv ../benchmark/
    cd ../../../pointcnn/data_conversions
    python -f ../../data/scannet/scannet_dataset_download/data/ -b ../../data/scannet/scannet_dataset_download/benchmark/ -o ../../data/scannet/cls/
    python -f ../../data/scannet/cls/
    cd ../pointcnn_cls/
    ./ -g 0 -x scannet_x3_l4
    • tu_berlin

    cd data_conversions
    python3 ./ -d tu_berlin
    python3 ./ -f ../../data/tu_berlin/ -a --create-train-test
    cd ../pointcnn_cls
    ./ -g 0 -x tu_berlin_x3_l4
    • quick_draw

    Note that the training/evaluation of quick_draw requires LARGE RAM, as we load all stokes into RAM and converting them into point cloud on-the-fly.

    cd data_conversions
    python3 ./ -d quick_draw
    cd ../pointcnn_cls
    ./ -g 0 -x quick_draw_full_x2_l6
    • MNIST

    cd data_conversions
    python3 ./ -d mnist
    python3 ./ -f ../../data/mnist
    cd ../pointcnn_cls
    ./ -g 0 -x mnist_x2_l4
    • CIFAR-10

    cd data_conversions
    python3 ./ -d cifar10
    python3 ./
    cd ../pointcnn_cls
    ./ -g 0 -x cifar10_x3_l4
  • Segmentation

    We use farthest point sampling (the implementation from PointNet++) in segmentation tasks. Compile FPS before the training/evaluation:

    cd sampling
    • ShapeNet

    cd data_conversions
    python3 ./ -d shapenet_partseg
    python3 ./ -f ../../data/shapenet_partseg
    cd ../pointcnn_seg
    ./ -g 0 -x shapenet_x8_2048_fps
    ./ -g 0 -x shapenet_x8_2048_fps -l ../../models/seg/pointcnn_seg_shapenet_x8_2048_fps_xxxx/ckpts/iter-xxxxx -r 10
    cd ../evaluation
    python3 -g ../../data/shapenet_partseg/test_label -p ../../data/shapenet_partseg/test_data_pred_10 -a
    • S3DIS

    Please refer to data_conversions for downloading S3DIS, then:

    cd data_conversions
    mv S3DIS_files/* ../../data/S3DIS/out_part_rgb/
    ./ -g 0 -x s3dis_x8_2048_fps -a 1
    ./ -g 0 -x s3dis_x8_2048_fps -a 1 -l ../../models/seg/s3dis_x8_2048_fps_xxxx/ckpts/iter-xxxxx -r 4
    cd ../evaluation
    python3 -d <path to *_pred.h5>

We use a hidden marker file to note when prepare is finished to avoid re-processing. This cache can be invalidated by deleting the markers.

Please notice that these command just for Area 1 (specified by -a 1 option) validation. Results on other Areas can be computed by iterating -a option.

  • ScanNet

Please refer to data_conversions for downloading ScanNet, then:

cd data_conversions
cd ../pointcnn_seg
./ -g 0 -x scannet_x8_2048_k8_fps
./ -g 0 -x scannet_x8_2048_k8_fps -l ../../models/seg/pointcnn_seg_scannet_x8_2048_k8_fps_xxxx/ckpts/iter-xxxxx -r 4
cd ../evaluation
python3 -d <path to *_pred.h5> -p <path to scannet_test.pickle>
  • Semantic3D

Please check the free disk space before start, about 900 GB will be required.

cd data_conversions
mkdir ../../data/semantic3d/filelists
cd ../pointcnn_seg
./ -g 0 -x semantic3d_x4_2048_fps
./ -g 0 -x semantic3d_x4_2048_fps -l <path to ckpt>
cd ../evaluation
python3 -d <path to *_pred.h5> -v <reduced or full>
  • Tensorboard

    If you want to monitor your train step, we recommend you use the following command
    cd <your path>/PointCNN
    tensorboard --logdir=../models/<seg/cls> <--port=6006>