Guides

One task per page, from a dataset on disk to a trained, tested and extended detector; each page gives the commands for every method it applies to.

• Configuration

  ---

  One YAML file is the whole experiment: inherit a shipped config, override, validate.

• Datasets

  ---

  GraspNet-1Billion and GraspClutter6D on disk, and the label assets each method reads.

• Train and test

  ---

  Train one of the six methods, score it on the benchmark, resume a run.

• Cross-dataset evaluation

  ---

  Test GraspNet-1Billion weights on GraspClutter6D, with only the test data changed.

• Your camera and Python

  ---

  Grasps on your own depth images, from the command line, Python or a browser.

• Read an analysis

  ---

  Where the AP of a result comes from, and whether two results really differ.

• Compose and probe

  ---

  Another option in a slot of a method, checked on real frames in a minute.

• Add a component

  ---

  Write, register and use your own encoder, stage, hook or evaluator without forking.