dobkeratops
commented
6 years ago EDIT: ok I have another workaround,
instead of searching for road&~sky (to add sky labels), i can search for ~sky&~tree etc (i.e. several common labels), then each time I click an image I have a reason to make a label choice through the existing interface (.. "this one has sky... this one has tree.. etc")
after using this for a while,
Seems like there might be scope for a variation of the browser mode where clicking adds (or removes) a label directly: (e.g. searching for&~sky usually finds more sky than not)
perhaps this could be useful for refinement too, e.g. search for dog confirm
1 dog, search forpersonand confirm cases ofman, etc .. so you'd need a way to specify a 'master label' again (the first? or the asterisk idea? or another entry box in a general purpose mode with an action dropdown?)This would probably need a line of text below each image showing it's status regarding the task, so you can see the click toggling it. (I've seen the popup text for showing the labels, thats very useful of course already) that might be a nice capability to have in the browse mode generally, e.g. when you search for a label, it could show other meta-information associated with the label perhaps
I think validation could be very fast in a similar mode too (because when you scroll through, anything different jumps out)
I dont think this is urgent, just something to think about. Another way to get safe training assumptions regarding common labels would be a confirmed absence, e.g.
no skyor~sky, although it might need a better name since we have "no entry sign", I recall ..no sky:for: reads naturally. against: disadvantage: a label like 'no entry sign' includes 'no'. (fixes.. use underscores or concatenation eg no_entry_sign , noentrysign? make the NO prefix uppercase?)~skyfor: similar to the search query. against: you also might want to search for the presence or absence of the 'no sky' label..other ideas.. displaying image ID's for cut/paste, and having an endpoint for script-driven advanced manipulations?
For reference think about the workflow of selecting image filenames via thumbnails in a filesystem.. this is ok, but the site has the opportunity of non-hierarchical grouping, and larger thumbnails, task focussed text information
seems for certain labels, e.g. 'apple', it's safe to assume a random image doesn't have it, whilst for common ones (like sky) it isn't. Perhaps comparing the number of instances of a label compared to the total number of labelled images would give an insight, or perhaps we need to identify these cases through common sense..
yet another idea would be to assume probability through co-occurence, i.e.
road,car,treegive a high probability of sky;sofa,ceiling,carpet,bedgive a low probability of sky