Skip to content

README.md

Latest commit

 

History

History
43 lines (37 loc) · 2.08 KB

README.md

File metadata and controls

43 lines (37 loc) · 2.08 KB

Scale-Shape-Grasp-Auto-Encoder

Instructions for data generation and training of scale-shape-grasp auto-encoder.

Data Generation

To generate this data:

  • Download NOCS dataset.
  • Configure various project level paths inside config_global_paths.py file. This include providing path to NOCS dataset root.
  • Run the following script to generate data:
python scene_grasp/scale_shape_grasp_ae/generate_grasp_nocs_scale_dset.py

Inside above script, each mesh goes through the following steps:

  • A random scale is applied to the mesh. This scale sampled from the scale distribution of the category. This scale distribution is obtained via ShapeNet-Sem dataset which has meshes in their real-world scale.
  • Grasp data is generated using our geometric grasp generation tool.
  • Finally, the data is processed and saved in pickle files that is easy to be used for network training.

Visualize the generated data

Use evaluate.py script which will generate the html visualizations of the generated data.

Training

For training, run the following command:

python scene_grasp/scale_shape_grasp_ae/train_ae_mine.py --dataset_root=<path-to-dataset-generated-above>

Visualize the trained network predictions

  • To visualize the shape and grasp success:
python scene_grasp/scale_shape_ae/evaluate_ae.py --dataset_root <path-to-dataset-root> --model_path <model-path>

This will generate HTML visualizations of predicted shape and grasp success at different scales and compare them against the ground-truths.

  • To visualize the grasp orientations in detail:
python scene_grasp/scale_shape_ae/evaluate_ae.py --dataset_root <path-to-dataset-root> --model_path <model-path>

This will iterate over datapoints one-by-one, showing the ground-truth and prediction one-by-one. This visualization will have grasp orientations and grasp width as well.