Get started¶
From install to your first swap in three steps. Nothing needs a dataset until you probe or train on real frames.
-
1 · Install
Install from source, then check it with a toy detector that trains in under a minute.
-
2 · Run a method
Pick a method and get grasps from its released weights, on the example scene or your camera.
-
3 · Swap a component
Replace an encoder or another stage with one line, check it, try it, train it.
-
Optional · Tour the config
See how a config is checked, and how a mistake is caught before anything is built.
Words used on these pages¶
| Word | What it means |
|---|---|
| slot | a stage of a detector you can replace: --set <slot>=… on the command line, model.<slot> in a config |
| option | what goes into a slot: a preset or a component name |
| seed-based skeleton | the pipeline five methods share: scene encoder → seeds → view → grouper → head; HGGD is its own family |
| scene encoder | the point-cloud encoder of the seed-based methods; its slot is backbone |
| image encoder | the image encoder of HGGD; its slot is image_encoder |
composes · runs · learns · predicts |
the levels a swap is measured at: config accepted · trains on real frames without error · its losses fall · it ranks grasps |
| probe | a short training on real frames that ends with one verdict, such as runs or refused; wired means a synthetic scene only |
| trains like shipped | a swap trained longer beside the component the method ships: on a par, worse, diverging, did not train or not comparable |
The exact definitions and thresholds: How we measure.