This repository represents the end-to-end pipeline of our multi-stage RGB-D image to point cloud completion architecture using two deep neural networks. It contains the following components:
- Web Component: RESTful web client for uploading YCB-Video data set frames and specifying the object you want to complete.
- VSN Component: Contains the customized Vanilla SegNet model for semantic segmentation and its dependencies.
- MSN Component: Contains the customized Morphing and Sampling Network for point cloud completion and its dependencies.
Each of the components is containerized using Docker, thus they are all self-contained so that no dependencies need to be installed other than Docker the ones listed below in the setup instructions.
Watch this to see the GUI in action. You can also find more details on the project here.
Beware that you need a GPU to run the pipeline!
- Download the Docker client for your OS.
If you are on a Linux system, make sure to add your user to the docker group as described here.
- The Morphing and Sampling Network can only be executed on a GPU, hence you will also need Nvidia-docker.
- Install Node.js
- Install yarn
- Clone this repository locally.
- Run
bash setup.sh
in the root of this repo to clone the VSN, MSN and Web Component repos and to create the necessary files on your machine automatically. Be patient, it takes a while initially :) By default, all the files will be created in the directory$HOME/Desktop/CVP
. If you want to change this, you can pass a different path using the-d
option when running setup.sh. - Download both network weight files from GDrive
and move them to
$HOME/Desktop/CVP/pipeline/models
or underpipeline/models
at the custom location you specified using-d
. - Run
bash ./run/start-web-app.sh
. Add the-d
option if you used it while running setup.sh. - In a new terminal window run
bash ./run/start-vsn.sh
. Add the-d
option if you used it while running setup.sh. - In a new terminal window run
bash ./run/start-msn.sh
. Add the-d
option if you used it while running setup.sh. - Navigate to http://0.0.0.0:5000 to see the Web Component's user interface.
- Pick the YCB object you want to create a completed point cloud for in the dropdown.
- Select files from the
$HOME/Desktop/CVP/pipeline/example_inputs
directory (or where you specified using-d
) corresponding to the object you picked (banana, bleach cleanser, power drill or scissors). Make sure to select all files belonging to the frame:<frame-id>-color.png
,<frame-id>-depth.png
and<frame-id>-meta.mat
. - Wait for the completion to finish. It will take 15+ seconds for both networks to process the inputs (depends on your GPU).
- Enjoy the resulting partial and completed point cloud displays directly in the browser. You can zoom in and rotate them within the display windows.
You can check out the other component repos on Gitlab.
- Vanilla Segnet: https://github.com/j96w/DenseFusion
- Morphing and Sampling Network: https://github.com/Colin97/MSN-Point-Cloud-Completion