Skip to content

classification.md

Latest commit

 

History

History
127 lines (108 loc) · 5.61 KB

classification.md

File metadata and controls

127 lines (108 loc) · 5.61 KB
Raw

Shape Classification

Change the working directory to caffe/experiments/cls for Caffe and tensorflow/scripts for TensorFlow.

O-CNN on Caffe

  1. Download ModelNet40 dataset and unzip it to the folder dataset/ModelNet40.

  2. Convert triangle meshes (in off format) to point clouds (in points format) with the virtual_scanner. This process can be automatically executed by the following command. Remember to provide actual <The path of the virtual_scanner> to run the command, and replace the symbol ^ with \ for multiple-line commands in the shell. We also provide the converted point clouds for convenience. Download the zip file here and unzip it to the folder dataset/ModelNet40.points.

    python prepare_dataset.py --run=m40_convert_mesh_to_points ^
                          --scanner=<The path of the virtual_scanner>

  3. The generated point clouds are very dense, and if you would like to save disk spaces, you can optionally run the following command to simplify the point cloud.

    python prepare_dataset.py --run=m40_simplify_points ^
                          --simplify_points=<The path of simplify_points>

  4. Convert the point clouds to octrees, then build the lmdb database used by caffe with the executive files octree and convert_octree_data. This process can be automatically executed by the following command. Remember to provide actual <The path of the octree> and <The path of the convert_octree_data>to run the command. We also provide the converted lmdb for convenience. Download the zip file here and unzip it to the folder dataset.

    python prepare_dataset.py --run=m40_generate_ocnn_lmdb ^
                          --octree=<The path of the octree> ^
                          --converter=<The path of the convert_octree_data>

  5. Run caffe to train the model. For detailed usage of caffe, please refer to the official tutorial Here. We also provide our pre-trained Caffe model in models/ocnn_M40_5.caffemodel. The classification accuracy on the testing dataset is 89.6% as reported in our paper. If the voting or view pooling operations are performed, the accuracy will be 90.4%

    caffe train  --solver=ocnn_m40_5_solver.prototxt  --gpu=0

AO-CNN on Caffe

  1. Follow the instructions above untile the 3rd step. The AO-CNN takes adaptive octrees for input, run the following command to prepare the data automatically. 
    python prepare_dataset.py --run=m40_generate_aocnn_lmdb ^
                          --octree=<The path of the octree> ^
                          --converter=<The path of the convert_octree_data>
  2. Run caffe to train the model. 
    caffe train  --solver=aocnn_m40_5_solver.prototxt  --gpu=0

O-CNN on TensorFlow

  1. Download ModelNet40 dataset and unzip it to the folder dataset/ModelNet40 via the following command:

    python ../data/cls_modelnet.py --run download_m40

  2. Convert triangle meshes (in off format) to point clouds (in points format) with the virtual_scanner. This process can be automatically executed by the following command. Remember to provide actual <The path of the virtual_scanner> to run the command.

    python ../data/cls_modelnet.py --run m40_convert_mesh_to_points \
                               --scanner <The path of the virtual_scanner>

    The generated point clouds are very dense, and if you would like to save disk spaces, you can optionally run the following command to simplify the point cloud.

    python ../data/cls_modelnet.py --run m40_simplify_points

    We also provide the converted point clouds for convenience. Download the zip file here and unzip it to the folder dataset/ModelNet40.points.

    python ../data/cls_modelnet.py --run download_m40_points

  3. The Tensorflow takes TFRecords as input, run the following command to convert the point clouds to octrees, then build the TFRecords database

    python ../data/cls_modelnet.py --run m40_generate_octree_tfrecords

  4. Run the following command to train the network. The performance is consistent with the Caffe-based implementation, i.e. the classification accuracy is 89.6% without voting.

    python run_cls.py --config configs/cls_octree.yaml

  5. With Tensorflow, the network can also directly consume the points as input and build octrees at runtime. Run the following command to store the points into one TFRecords database.

    python ../data/cls_modelnet.py --run m40_generate_points_tfrecords

  6. Run the following command to train a deeper network with ResBlocks, which directly takes points. Notable, simply using the training hyperparameters as before, the testing accuracy increases from 89.6% to 92.4%.

    python run_cls.py --config configs/cls_points.yaml