ShapePFCN

Shape Projective Fully Convolutional Network

This is the implementation of the ShapePFCN architecture described in this paper:

Evangelos Kalogerakis, Melinos Averkiou, Subhransu Maji, Siddhartha Chaudhuri, "3D Shape Segmentation with Projective Convolutional Networks", Proceedings of the IEEE Computer Vision and Pattern Recognition (CVPR) 2017 (oral presentation)

Project page: http://people.cs.umass.edu/~kalo/papers/shapepfcn/index.html

Arxiv most recent version: https://arxiv.org/abs/1612.02808

---

To compile in Linux (we assume 32 threads for compilation, change make's -j32 option according to your system):

1. First compile Siddhartha Chaudhuri's "Thea" library:

• compile Thea's dependencies:

     cd TheaDepsUnix/Source/
     ./install-defaults.sh --user <your_user_name> --with-osmesa -j32
     cd ../../
     cp -R TheaDepsUnix/Source/Installations/include/GL TheaDepsUnix/Source/Mesa/mesa-11.0.7/include

• compile the Thea library (note: adjust the path to the cuda directory according to your system)

     cd Thea/Code/Build
     cmake -DTHEA_INSTALLATIONS_ROOT=../../../TheaDepsUnix/Source/Installations/ -DTHEA_GL_OSMESA=TRUE -DOSMesa_INCLUDE_DIR=../../../TheaDepsUnix/Source/Mesa/mesa-11.0.7/include/ -DOSMesa_GLU_LIBRARIES=../../../TheaDepsUnix/Source/Mesa/mesa-11.0.7/lib -DOPENCL_INCLUDE_DIRS=/usr/local/cuda75/toolkit/7.5.18/include  -DOPENCL_LIBRARIES=/usr/local/cuda75/toolkit/7.5.18/lib64 -DCMAKE_BUILD_TYPE=Release
     make -j32
     cd ../../../
     ln -s Thea/Code/Build/Output/lib lib

2. Given that Thea's libraries were compiled successfully (for questions related to Thea, please email Siddhartha Chaudhuri), the next step is to compile our version of caffe (sorry, we modified caffe to incorporate our own data & projection layers) and generate the header used for parsing protobuf schemas in Caffe:

     cd caffe-ours   
     make -j32
     sh generate_proto.sh     
     cd ../

(notes: you may need to adjust the library paths in `caffe-ours/Makefile.config' according to your system, and you also need to install the libraries that caffe requires: http://caffe.berkeleyvision.org/installation.html)

3. Given that caffe was compiled successfully, you can now compile ShapePFCN. In the root directory of ShapePFCN, type:

     make -j32

(note: you may need to adjust the library paths in Makefile.config according to your system)

4. Download the pretrained VGG model on ImageNet from here : http://neghvar.cs.umass.edu/public_data/vgg_conv.caffemodel (we train starting from a pretrained VGG model). Place it in the ShapePFCN root directory (i.e., frontend_vgg_train_net.txt and vgg_conv.caffemodel should be in the same directory)

---

To run the net training procedure (the first command renders images, the second runs the network training):

     ./build_release/mvfcn.bin --skip-testing --do-only-rendering --train-meshes-path  <your_path_to_training_data>
     ./build_release/mvfcn.bin --skip-testing --skip-train-rendering --train-meshes-path  <your_path_to_training_data> --gpu-use 0

Notes:

• the gpu-use option specifies the id of the GPU to use in your system, 0 means your first GPU card)
• for more options, type: ./build_release/mvfcn.bin --help
• the implementation assumes a GPU card with 24GB memory (e.g., Tesla M40, Quadro P6000). If you don't have such GPU cards, add the following two arguments for both the above commands: '--pretraining-batch-splits 4' (for TitanX, 12GB memory) or '--pretraining-batch-splits 8' (for cards with <=8GB mem) and '--skip-mvfcn'
• if your shapes have consistent upright (gravity) orientation, PLEASE use the following arguments for both the above commands: --use-upright-coord and --up-vector "0.0 1.0 0.0" (if y-axis is gravity axis / change according to the upright axis of your dataset and use double quotes) (for more discussion about the benefits of consistent upright orientation, check the arxiv v3 version of our paper)
• for faster training, you may consider the option: '--baseline-rendering' for both the above commands, which renders models according to a fixed dodecahedron-based camera setting (the performance drop is minor)
• after rendering is completed (i.e., execution of the first command), there might be a crash at the very end of the program in some systems (after all the rendered images are produced). This might be due to a double deallocation of GL resources, which we unfortunately did not figure out why it happens. However, since this is done after all rendering is completed (check if all rendered images are produced for all your meshes), the training (and similarly testing) will not be affected. We aplogogize for the ``academic'' nature of the code - it is not bulletproof!
• you may want to adjust your LD_LIBRARY_PATH so that all required libraries are accessible e.g., in one of the systems we tried our code, before you run the above commands, we execute:

     LD_LIBRARY_PATH=$LD_LIBRARY_PATH:./caffe-ours/build/lib:/usr/local/hdf5_18/1.8.17/lib:/usr/local/openblas/0.2.18/lib:/usr/local/boost/lib:/usr/local/cuda75/toolkit/7.5.18/lib64:/usr/local/cudnn/5.1/lib64/:/usr/local/apps/cuda-driver/libs/375.20/lib64/
     export LD_LIBRARY_PATH

---

To run the testing procedure (after you execute training!):

     ./build_release/mvfcn.bin --skip-training --do-only-rendering --test-meshes-path  <your_path_to_test_data> --train-meshes-path <your_path_to_training_data>
     ./build_release/mvfcn.bin --skip-training --skip-test-rendering --test-meshes-path  <your_path_to_test_data> --train-meshes-path <your_path_to_training_data> --gpu-use 0

Same notes as above apply wrt GPU usage, memory, shape orientation, and "baseline" rendering.

For any questions related to the compilation and execution of ShapePFCN and our caffe version, you may contact Evangelos Kalogerakis

---

Regarding training/test data format:

Our repository includes the airplanes from the L-PSB dataset (http://people.cs.umass.edu/~kalo/papers/LabelMeshes/index.html) as an example of the data format that ShapePFCN supports. There are two possible formats:

• OBJ files where each training part is stored as a group in an OBJ file (see psbAirplane1 folder and https://en.wikipedia.org/wiki/Wavefront_.obj_file)
• OFF files where labels are stored in separate label txt files (see psbAirplane2 folder). In this case, each OFF mesh has a txt file that ends with the string "_labels.txt". Each pair of lines in these txt files contains a label identifier and a list of integers that are indices to faces having that label (the first face has index 1, not 0).

For testing, no OBJ groups or labels txt files are needed. If they are found in the test directory, they will be simply used for evaluating test accuracy.