voxel_yolonet

Reproducing VoxelNet from Apple, forked from https://github.com/jeasinema/VoxelNet-tensorflow Adding YOLO-structure

SETTING

Preparing for final upload

Log: Aug 13, 2018, Monday

Requirements may include:

python 3.5
tensorflow 1.4
numpy
opencv

Folder structure

[root]
 data       : kitti dataset( : should be linked to your original kitti dataset)
 model      : model
 utils      : utils files
 
 and some files such as data_aug, evaluate_object, misc_utils, train .. are not located inside folders

Structure of data folder

data -> object -> [training, testing, ...] training -> [calib, image_2, label_2, velodyne, ..] testing -> [calib, image_2, velodyne, ..] <- no label_2 folder in testing folder

RUN

train

All configurations are defined in config.py

cd [root]
CUDA_VISIBLE_DEVICES=2,3 python3.5 train.py

To use multiple gpus, you need to modify __C.GPU_AVAILABLE='0,1' in config.py file If you use 3 gpus, __C.GPU_AVAILABLE='0,1,2' regardless of the actual numbering of your gpus such as (gpu 0, gpu 2, gpu 3) in CUDUA_VISIBLE_DEVICES=0,2,3

It will save model in 'save model' folder that will be automatically generated. It also save images in 'save_image' folder which also be generated automatically. And finally, it calls validation program to save the result in result folder and plot folder inside in it.

test

Currently no test script. It is included in validation step.

Kiwoo / voxel_yolonet