Rec4Aug: An Efficient Pipeline for Traffic Scene Data Generation

🧑‍🎓 Members:

HanzhiC
Spazierganger
DaoyiG

📃 Project Description:

Our pipeline first collects a sequence of monocular RGB images and predicted depth for large-scale dense 3D scene reconstruction, and then automatically place virtual objects on top of ground in order to generate realistic traffic scene data (manual mode is enabled as well for fine tunning or generating 3D occlusion awareness).

1. Depth Prediction

2. Reconstruction

3. Augmentation

Auto Mode:

Manual Mode:

⚙️ Project Environment:

Create our environment with

conda create -n rec4aug python=3.6.6

After activate the environment rec4aug, you need to install:

conda install pytorch=0.4.1 torchvision=0.2.1 -c pytorch
pip install open3d
pip install opencv-python

🧰 Additional Dependencies:

We use imagemagick to transform the output of monodepth to the format taken by InfiniTAM

🎬 Run Whole Pipeline

You can simply use terminal under this directory and type

bash run_pipeline.sh

to run the whole pipeline of our project. You can choose whether to skip a stage( e.g. reconstruction using open3d) by entering y or N with respect to corresponding shell prompt.

🔗 Related Works:

Monodepth2

InfiniTAM v3

Open3d Reconstruction Pipeline