Image2PCL

Enter the metaverse with 2D image to 3D projections!
This is an implementation of an algorithm to project 2D images into the 3D space. See below for a visual summary of the project

Setup

Assuming you have already set up an Anaconda environment with PyTorch, CUDA and Python, install additional dependencies with:

pip install open3d
pip install mmcv-full=={mmcv_version} -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10.0/index.html

Clone the mmsegmentation repository to your working directory

git clone https://github.com/open-mmlab/mmsegmentation

Create a 'models' folder to store your trained models for testing.

mkdir models

Trained KITTI models can be downloaded from the monodepth2 repository. This code was tested with the 'mono_640x192' model.
I also provide a custom-trained nuScenes model for testing with nuScenes images. This is helpful for multi-view point cloud rendering.

Test

To run a test, it is preferred to use images from a dataset with known camera intrinsics. For this implementation, we use two different datasets:

• KITTI Raw for single image testing
• nuScenes for multi-view images testing

To test on KITTI, run the following (replace the "<>" brackets and contents inside with the correct information):

python img2pcl.py \
--image_path <path to single image file or folder containing single image> \
--model_path <path to trained KITTI model> \
--data_type kitti_raw

To test on nuScenes to view a 360 3D point cloud, run the following (replace the "<>" brackets and contents inside with the correct information):

python img2pcl.py \
--image_path <path to folder containing nuScenes multi-cam images> \
--model_path <path to trained nuScenes model> \
--data_type nuscenes \
--nusc_camera_parameters <path to a json file containing nuscenes camera intrinsics and extrinsics>