search

ImageMonkey / imagemonkey-core

ImageMonkey is an attempt to create a free, public open source image dataset.
https://imagemonkey.io
47 stars 10 forks source link

brainstorming - anti-labels for area coverage (and alternative to annotating all) #152

Open dobkeratops opened 6 years ago

dobkeratops commented 6 years ago

so when you add a label to a scene, the pixels are 'maybe' imagine if you could work at the problem from 2 ends, adding the pixels that definitely are, and removing the pixels that definitely aren't - the latter being easier for dense objects,(e.g. trees in a forest)

here's an example of what it could look like e.g. the red area being "not tree"

screen shot 2018-06-25 at 15 13 33

imagine if there was always the ability to toggle into a negative mode ('not tree') , ('not traffic', 'not building..')

A verified "all" annotation (imagemonkey's current default) could be considered a polygon who's inverse is automatically also the 'not label'.

But the advantage of having both would be leaving an indeterminate area in the middle for fuzzy objects ,e.g. faint branches, patchy grass, etc.. you could imagine generating feathered mask by filling an area with '0.5', then setting the posative and negative labels to 1.0 and 0.0 respectively, then blurring the 0.5 area between them.

dobkeratops commented 6 years ago

( I'm sort of already counting on occlusion rules to do some clipping, e.g. assuming 'objects' like car, person,dog/cat will occlude the anotation of 'ground' types like road,pavement,grass, river,sand.. )