The release includes source code for OS-TOG, model weights, the database used for real-world experiments, and manually annotated object masks for the "UMD RGB-D Part Affordance Clutter" dataset.
Model Weights
grasp_detection_model.pt- model weights for the grasp detection model, trained on the Cornell grasping dataset.instance_segmentation_model.pt- model weights for the instance segmentation model, trained on the OCID, UMD tools, and UMD clutter dataset.object_recognition_model.pt- model weights for the object recognition model, trained on the UMD tools dataset.
Data and Annotations
UMD_clutter_annotations.pt- JSON annotation file that has object mask annotations for 30 scenes from the UMD clutter dataset that were manually annotated and used to train the instance segmentation model.OSTOG_physical_experiments.json- JSON annotation file of the objects and tasks the system knows that was used in real-world experiments.OSTOG_physical_experiments.zip- images of reference objects and their annotated affordance for the database created for real-world experiments.