Invisible shape identification #60
Conversation
src/penai/svg.py
Outdated
| @@ -538,6 +566,20 @@ def is_container_type(self) -> bool: | |||
| def is_primitive_type(self) -> bool: | |||
| return self._shape_type.value.category == PenpotShapeTypeCategory.PRIMITIVE | |||
|
|
|||
| @property | |||
| def produces_visible_content(self) -> bool: | |||
There was a problem hiding this comment.
can we turn this into a method pls? I expect properties access to be essentially instantaneous (like attr access)
|
@MischaPanch After no progress last week, I have now updated the code and fixed the outstanding issues. You can merge it if you are happy with the current state. |
src/penai/xml.py
Outdated
| if key.startswith("ns"): | ||
| from lxml import etree | ||
|
|
||
| print(etree.tostring(self)) |
There was a problem hiding this comment.
pls no printing in properties (or in fact almost anywhere). Did this whole block sneak in accidentally during debugging?
There was a problem hiding this comment.
I should stop doing debug sessions after midnight...
|
@kklemon LGTM! Just minor comments about something that looks like forgotten/debugging code. Feel free to merge yourself after fixing |
test/penai/test_svg.py
Outdated
| hook, | ||
| ): | ||
| # Even though the underlying SVGs have the perfectly same visible content, they still | ||
| # might produc |
There was a problem hiding this comment.
finish sentence or remove ;)
Closes #52.
While this was initially meant as a simple-to-implement improvement for the bounding box generation for the VLMs, it turned out to be much more difficult to implement and particularly to test than anticipated.
The few simple rules that were sketched in #52 were not sufficient and consideration of many different edge cases and interactions between various SVG element attributes was necessary.
In the end, I tried to keep the logic as simple as possible with the aim to reduce the false positive rate while possibly missing a few invisible shapes. The provided tests do in fact test for pixel perfectness between the original and cleaned version and and thus should guarantee a FPR of 0, given that the tests are correct.
Here is the effect visualized with bounding boxes:
TODO: