A tensorflow implementation for VoxelNet.
Python 3.5+tensorflow 1.4+NumPy, etc.have a look at config.py for model configurations, split your data into test/train set by this.
run setup.py to build the Cython module.
$ python setup.py build_ext --inplacemake sure your working directory looks like this (some files are omitted):
├── build <-- Cython build file
├── model <-- some src files
├── utils <-- some src files
├── setup.py
├── config.py
├── test.py
├── train.py
├── train_hook.py
├── README.md
└── data <-- KITTI data directory
└── object
├── training <-- training data
| ├── image_2
| ├── label_2
| └── velodyne
└── testing <--- testing data
├── image_2
├── label_2
└── velodyne run train.py. Some cmdline parameters is needed, just check train.py for them.
launch a tensorboard and wait for the training result.
Since c928317, data augmentation is done in an online manner, so there is no need for generating augmented samples.
TBD
Thanks to @ring00 for the implementation of VFE layer and Jialin Zhao for the implementation of the RPN.
MIT