Repository files navigation

This project implementsDeep Closest Point model in TensorFlow. It also includes C++ code that compare its performance with other registration methods (ICP, 4-PCS, Go-ICP).

To run DCP model, you may have to install these Python packages:

• tensorflow>=2.0.0
• tensorflow-graphics (none of its dependencies is required)
• numpy
• h5py

To run comparison program, you may have to install these libraries:

• PCL 1.9 (and its dependencies)
• HDF5
• TBB

Basic usage is encapsulated into procedures. You can directly call them in the program. Hyperparameters are directly defined in source code, and command line arguments is not supported.

DownloadModelNet40 and unzip files into directorymodelnet40. Runutil.pack_to_one() to pack all dataset files into singletrain.h5 andtest.h5 files.

Trained weightsdcp_v2.h5 can be unzipped fromweights/dcp_v2.zip. Place it inweights directory so that evaluation and testing procedure can find it. If you want to train by yourself, runtrain.train() to train, or your owning training procedure. Runtrain.evaluate() to evaluate the trained model with test dataset.

The comparison program tests registration methods on the first 100 models of the test dataset. It is divided into Python and C++ code. Runcompare.test_dcp() to test DCP. Compile and run the C++ program to test ICP, 4-PCS and Go-ICP. ICP and 4-PCS implementation is from PCL. Go-ICP is from my previous projectOptICP.

The projectproposal andreport are provided (both in Chinese). Refer to them for better understanding of this project.