Autoscale geometries

[changhiskhan]

Currently if we have a relative Box, we need to manually rescale it by the image in order to draw it:

img | [box * (row.height, row.width) for box in row.annotations]

Instead we should consider:

1. Indicate whether a Box or Polygon is in relative or absolute space.
2. Add a property in Image to retrieve the shape
3. Automatically rescale the geometry when drawing

[changhiskhan]

@Renkai if you're looking for something to tackle, this is a good one

[changhiskhan]

One alternative to adding an attribute is we could consider creating a RelativeGeometry mixin that has functionality to autoscale. Then we just create Relative* subclasses of * Geometry classes with the RelativeGeometry mixin added and (hopefully) nothing else. Thoughts? @eddyxu ?

[Renkai]

@changhiskhan Sounds interesting, but I have some questions.

• What is Add a property in Image to retrieve the shape?
• What is the mixin here? In Scala, I think it's trait, but in Python, I'm not sure what is a proper way to implement mixin.

And can you address some links to the key classes or files on how we implement the box drawing mechanism in the current code base?

[da-liii]

Key python file: https://github.com/eto-ai/rikai/blob/main/python/rikai/viz.py

[da-liii]

Mixin: https://github.com/eto-ai/rikai/blob/main/python/rikai/mixin.py

https://github.com/eto-ai/rikai/blob/fed57cd7d22b520b87c1a6693850916051ba9331/python/rikai/types/vision.py#L44

Rikai Image depends on Mixins heavily.

[da-liii]

Here is a related issue: https://github.com/eto-ai/rikai/issues/523

These DSL should be designed at the very beginning. And then we start to implement it.

[da-liii]

I found that the box2d returned by AWS Rekognition (e.g. recognize celebrities) is relative box2d.