(see also #261) So..
many words that could be properties make sense for many completely different object types
But some have visual similarities, whilst others dont.
example - derelict, vs asian
"derelict" has visual cues of dirt, holes,rust, broken uneven shapes
a derelict house, derelict car, derelict* aeroplane all have a bit in common visually with eachother (such that if you trained on those, you could get a "derelict" visual property to guess for other objects.. like independent recognition of textures that we mentioned)
However the nationality or region varitations probably don't
e.g. there's no visual simiarity between asian person and asian text - the only link through asian is that both things would appear more in photos from china,japan etc.
I mention this because it just occurred to me that this confuses the decision to use a "/" seperator.
Not sure what the answer is here.
When using "/" the implication is sometimes that A/B would light up an output for both words A and B independently.
e.g. we've already said how part names - wheels, heads etc - probably do have a lot in common with each other between completely different object classses. Even a "leg/table" and "leg/giraffe" have something in common - they're vertical supporting structures.
So my initial instinct was to want to write "asian/text", "english/text" . However in light of the above, you might want to combine the nationality or region property more tightly with the object?
any thoughts on this?
I guess it might not be a problem, but it might potentially how you want to organise and train properties.
(i'm kind of assuming some words for properties will be combinable via "/" , the "derelict" example above, and some materials, pending awkward ambiguity like 'glass')
Maybe (contrary to my concern) the visual co-occurence (within "photos from china" etc) could actually make sense, and as such making a guess "its from asia" because a certain object appeared more in the same photos as an "asian/person" or "asian/text" could actually be valid?
(see also #261) So.. many words that could be properties make sense for many completely different object types
But some have visual similarities, whilst others dont.
example - derelict, vs asian
"derelict" has visual cues of dirt, holes,rust, broken uneven shapes a derelict house, derelict car, derelict* aeroplane all have a bit in common visually with eachother (such that if you trained on those, you could get a "derelict" visual property to guess for other objects.. like independent recognition of textures that we mentioned)
However the nationality or region varitations probably don't e.g. there's no visual simiarity between asian person and asian text - the only link through asian is that both things would appear more in photos from china,japan etc.
I mention this because it just occurred to me that this confuses the decision to use a "/" seperator.
Not sure what the answer is here.
When using "/" the implication is sometimes that A/B would light up an output for both words A and B independently. e.g. we've already said how part names - wheels, heads etc - probably do have a lot in common with each other between completely different object classses. Even a "leg/table" and "leg/giraffe" have something in common - they're vertical supporting structures.
So my initial instinct was to want to write "asian/text", "english/text" . However in light of the above, you might want to combine the nationality or region property more tightly with the object?
any thoughts on this?
I guess it might not be a problem, but it might potentially how you want to organise and train properties. (i'm kind of assuming some words for properties will be combinable via "/" , the "derelict" example above, and some materials, pending awkward ambiguity like 'glass')
Maybe (contrary to my concern) the visual co-occurence (within "photos from china" etc) could actually make sense, and as such making a guess "its from asia" because a certain object appeared more in the same photos as an "asian/person" or "asian/text" could actually be valid?