Skip to content

Latest commit

 

History

History
85 lines (67 loc) · 3.44 KB

README.md

File metadata and controls

85 lines (67 loc) · 3.44 KB
Raw

Experiments

cd ./  # root directory of the cloned repository
conda activate CompNet

Data

Download the following data and unzip it to data/.

  1. Train
  2. Test with ground truth part masks
  3. Test with predicted part masks

Pre-trained CompNet

Download the pretrained models and unzip to ./.

The above link contains pretrained models including rotation module, axis length module, joint module.

Testing CompNet

To reproduce the prediction results of the paper, run the following code

  • Predict shapes with ground truth part masks
bash ./scripts/predict_shape_with_gt_mask.sh chair
bash ./scripts/predict_shape_with_gt_mask.sh bed
bash ./scripts/predict_shape_with_gt_mask.sh table
bash ./scripts/predict_shape_with_gt_mask.sh cabinet
  • Predict shapes with ground truth part masks
bash ./scripts/predict_shape_with_predicted_mask.sh chair
bash ./scripts/predict_shape_with_predicted_mask.sh bed
bash ./scripts/predict_shape_with_predicted_mask.sh table
bash ./scripts/predict_shape_with_predicted_mask.sh cabinet

Training SDFNet

Rotation module

python ./CompNet/train.py --cfg ./configs/RotNet.yaml

Axis length module

We provide two choices of axis length module: unary axis length module and group axis length module.

  • Train unary axis length module

python ./CompNet/train.py --cfg ./configs/AxisLengthNet.yaml

  • Train group axis length module. Two seperate modules need to be trained. One classify whether two parts are translation symmetry and another module predict axis length given a group of parts.
python ./CompNet/train.py --cfg ./configs/GroupAxisLengthNet.yaml 
python ./CompNet/train.py --cfg ./configs/SizeRelationNet.yaml

In experiments, unary axis length module also generates satisfactory predictions.

Joint module

python ./CompNet/train.py --cfg ./configs/JointNet.yaml

Tools

We provide visualization tools to help understand each module predictions.

  • After run the testing code bash ./scripts/predict_shape_with_gt_mask.sh bed
  • Go the prediction directory cd ./outputs/prediction_gt_mask/bed
  • Run python -m http.server 8003 and open link

The link provides prediction and visualization as follows.

    rot/                   # contains prediction results after rotation module
        table.html         # collections of predictions
        ...    
    size/                   # contains prediction results after rotation module and axis length module
        table.html
    center/                  # contains the final prediction results
    　　table_seq.html

Illustration of table.html. VIS_TABLE

Example visualization of rotation module. Note the visualization below takes ground truth size and center. VIS_ROT

Example visualization of size module. Note the visualization below takes ground truth center. VIS_SIZE

Example visualization of center module. Note we sequentially assemble the shape together. All rotation, size, and center are from predictions. VIS_CENTER