• Python 3.5.2 or higher
• TensorFlow-GPU 1.3.0 or higher
  • CUDA
  • cuDNN
  • GPU

First, change -arch=sm_52 in the files lib/setup.py and lib/roi_pooling_layer/make.sh to an appropriate value for your GPU.

./make.sh

Download left color images, Velodyne point clouds, camera calibration matrices, and training labels from https://www.cvlibs.net/datasets/kitti/eval_object.php. Split them into a training set and a validation set (e.g. according to Train/Val Split in https://www.cs.toronto.edu/objprop3d/downloads.php) and specify the directories of them by DIR_TRAIN and DIR_VAL in train.py.

Please read help by python train.py -h.

Please read help by python infer.py -h.

cd path/to/inference/output_dir
mkdir data
mv *.txt data/

cd anywhere/you/like
git clone https://github.com/prclibo/kitti_eval.git
cd kitti_eval
c++ -o evaluate_object_3d_offline evaluate_object_3d_offline.cpp
./evaluate_object_3d_offline path/to/validation/labels path/to/inference/output_dir