PON

Part Based Image To Structure

File Structure

/net ---network structure
/config ---configuration files
/util ---utility functions
/scripts ---scripts to run tasks as: train eval debug data-process /data ---path or link for data
/doc ---project documents
/docs ---host project home pages
/log ---log file path
run.py ---global main function, mainly define the running options

Dataset

the dataset for training and testing can be downloaded here

Environment Set Up

To run code from this repo, we recommend to use the docker image samhu/pytorch:latest from docker hub .

In addition, the chamfer distance used in this project need to be installed

as

pip install chamferdist

Training & Testing

The pipeline in this repo contains three networks that needed to be independently trained.

Training ``Touch Net", this network is used to tell if two parts are connected in 3D connected in 3D space

python run.py -X util.trainvalcage -net Touch -config Touch -bs 64 -ds CageNet -dp </path to dataset> -nepoch 50 -lrd 10 -lrdr 0.5 -key val -md full -rate 0.3

Training ``Box Net", this network is used to generate bounding box for each part

python run.py -X util.trainvalcage -net Box -config BoxBcd -bs 64 -ds CageNetTouch -dp </path to dataset> -nepoch 50 -lrd 10 -lrdr 0.5 -key val -md full -rate 0.3

Training ``Touch Point Net", this network is used to predict the relative translation in 3D space for two

python run.py -X util.trainvalcage -net TouchPt -config TouchPt -bs 64 -ds CageNetTouch -dp </path to dataset> -nepoch 50 -lrd 10 -lrdr 0.5 -key val -md full -rate 0.3

Testing

Result Preview

<iframe src=http://171.67.77.236:8082/bv/data~res~new_Chair_full~_6e1dd008531f95fc707cdefe012d0353_r60> </iframe>