/MV3D

Multi-View 3D Object Detection Network for Autonomous Driving

Primary LanguageJupyter Notebook

To clone,

$git clone --recursive  https://github.com/bostondiditeam/MV3D.git

To pull, use

$git pull --recurse-submodules

Contents

  • Key facts
  • Workflow
  • How to run
  • Todo
  • Issues

Key facts

Workflow

Please refer to here

Key Dependency

  • A Nvidia GPU card with computation capability > 3
  • CUDA
  • Python3.5 for MV3D related code
  • Tensorflow-GPU(version>1.0)
  • Python2.7 for ROS related script

File Structure

├── data   <-- all data is stored here. (Introduced in detail below)
│   ├── predicted  <-- after prediction, results will be saved here.
│   ├── preprocessed   <-- MV3D net will take inputs from here(after data.py) 
│   └── raw <-- raw data
├── environment_cpu.yml  <-- install cpu version.
├── README.md
├── saved_model                 <--- model and weights saved here. 
├── src        <-- MV3D net related source code 
│   ├── config.py
│   ├── data.py
│   ├── didi_data
│   ├── kitti_data
│   ├── lidar_data_preprocess
│   ├── make.sh
│   ├── model.py
│   ├── mv3d_net.py
│   ├── net
│   ├── play_demo.ipynb
│   ├── __pycache__
│   ├── tracking.py   <--- prediction after training. 
│   ├── tracklets
│   └── train.py    <--- training the whole network. 
│── utils    <-- all related tools put here, like ros bag data into kitti format
│    └── bag_to_kitti  <--- Take lidar value from ROS bag and save it as bin files.
└── external_models    <-- use as a submodule, basically code from other repos.
    └── didi-competition  <--- Code from Udacity's challenge repo with slightly modification, sync with Udacity's new
     updates regularly.

Related data are organized in this way. (Under /data directory)

├── predicted <-- after prediction, results will be saved here.
│   ├── didi <-- when didi dataset is used, the results will be put here
│   └── kitti <-- When kitti dataset used for prediction, put the results here
│       ├── iou_per_obj.csv   <-- What will be evaluated for this competition, IoU score
│       ├── pr_per_iou.csv   <--precision and recall rate per iou, currently not be evaluated by didi's rule
│       └── tracklet_labels_pred.xml  <-- Tracklet generated from prediction pipeline. 
├── preprocessed  <-- Data will be fed into MV3D net (After processed by data.py)
│   ├── didi <-- When didi dataset is processed, save it here
│   └── kitti <-- When Kitti dataset is processed, save it here
│       ├── gt_boxes3d
│           └── 2011_09_26
│               └── 0005
|                   |___ 00000.npy
├       |── gt_labels
│           └── 2011_09_26
│               └── 0005 
|                   |___ 00000.npy
|       ├── rgb
│           └── 2011_09_26
│               └── 0005 
|                   |___ 00000.png
|       ├── top
│           └── 2011_09_26
│               └── 0005 
|                   |___ 00000.npy
|       └── top_image
|           └── 2011_09_26
|               └── 0005 
|                   |___ 00000.png
└── raw  <-- this strictly follow KITTI raw data file format, while seperated into didi and kitti dataset. 
    ├── didi <-- will be something similar to kitti raw data format below. 
    └── kitti
        └── 2011_09_26
            ├── 2011_09_26_drive_0005_sync
            │   ├── image_02
            │   │   ├── data
            │   │   │   └── 0000000000.png
            │   │   └── timestamps.txt
            │   ├── tracklet_labels.xml
            │   └── velodyne_points
            │       ├── data
            │       │   └── 0000000000.bin
            │       ├── timestamps_end.txt
            │       ├── timestamps_start.txt
            │       └── timestamps.txt
            ├── calib_cam_to_cam.txt
            ├── calib_imu_to_velo.txt
            └── calib_velo_to_cam.txt

Modification needed to run

After Tensorflow-GPU could work If you are not using Nvidia K520 GPU, you need to change "arch=sm_30" to other value in src/net/lib/setup.py and src/lib/make.sh in order to compiler *.so file right. Here is short list for arch values for different architecture.

# Which CUDA capabilities do we want to pre-build for?
# https://developer.nvidia.com/cuda-gpus
#   Compute/shader model   Cards
#   6.1		      P4, P40, Titan X so CUDA_MODEL = 61
#   6.0                    P100 so CUDA_MODEL = 60
#   5.2                    M40
#   3.7                    K80
#   3.5                    K40, K20
#   3.0                    K10, Grid K520 (AWS G2)
#   Other Nvidia shader models should work, but they will require extra startup
#   time as the code is pre-optimized for them.
CUDA_MODELS=30 35 37 52 60 61

Test your Tensorflow-GPU is running by"

import tensorflow as tf
sess = tf.Session()
print(tf.__version__) # version more than v1.

It runs without error message and show ＂successfully opened CUDA library libcublas.so.8.0 locally＂, then it is in CUDA successfully.

cd src
source activate didi
sudo chmod 755 ./make.sh
./make.sh
# prerequisite for next step, i.e. running preprocessing using data.py, is to 
# follow steps in utils/bag_to_kitti if using didi data
python data.py # for process raw data to input network input format
python train.py # training the network.

Some other readme.md files inside this repo:

Issue

  • Not related to this repo, but if you are using Amazon CarND AWS AMI (Ubuntu 16.04 and with tensorflow-gpu 0.12 installed), pip install --upgrade tensorflow won't work and will introduce driver/software conflict. Because CarND AMI has a nvidia 367 driver, but after running above line, it will install 375 driver. I think in this case, tensorflow-gpu (version >1.0) need to compiled from source code.
  • If you already have a Tensorflow-GPU > 1, then the above ./make.sh works.
  • If you see error message "tensorflow.python.framework.errors_impl.NotFoundError: YOUR_FOLDER/roi_pooling.so: undefined symbol: ZN10tensorflow7strings6StrCatB5cxx11ERKNS0_8AlphaNumES3", it is related to compilation of roi_pooling layer. A simple fix will be changing "GLIBCXX_USE_CXX11_ABI=1" to "GLIBCXX_USE_CXX11_ABI=0" in "src/net/lib/make.sh" (line 17)