ImageMonkey / imagemonkey-core

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

scene label ideas #232

Open dobkeratops opened 5 years ago

dobkeratops commented 5 years ago

with the current upload page behaviour, if we had these , casual contribution would be easier: you could make an accurate statement at the point of upload. An accurate label for the whole image would give shortcut hints, and be useful for searches

street the most common catch all for urban scenes .. road+pavement+buildings+sky, very likely car+person+tree aswell, then some incidental objects - bus shelter telephone box litter bin ...

IMG_0718 IMG_0719

alleyway - very similar to street, but thinner, much less busy, maybe also 'backstreet'. more likely to have litter, bin

IMG_0717

construction site these usually have loads of interesting objects IMG_0623

IMG_0496 IMG_0494 IMG_0554

maybe rural path - implication of fields etc, sparse buildings IMG_0284

canal/marina .. not sure of the exact classification here.. there's sluice gates like a canal, there's an enclosed area for boats.. there is a canal connected to it - hence sometimes canal boats around here. ships dock at a port, move cargo onto canal boats?

IMG_0537

farm - fields, tractors, hay bales, sparse buildings.. more specific than rural road/path IMG_0132

IMG_0270 IMG_0250

park IMG_0633

park gardens .. 'gardens is more like 'public gardens', like a park.. sliding scale garden (singular) would be more like the lawn behind someone's house.

IMG_0586 IMG_0574

docks/marina IMG_0424

IMG_0447

park road

IMG_0636

riverside or waterfront

IMG_0558 IMG_0546

IMG_0318

junkyard

IMG_0514 IMG_0516

roadworks

IMG_0734

cathedral interior

IMG_0674

seaside - sea , beach, sky, maybe some buildings/trees, promenade, piers, maybe boats, probably people

IMG_0214 IMG_0082

IMG_0210

landscape catch all for areas predominantly natural (hills, beach, river, waterfall, mountains,valleys, lakes). in the case of these images, I couldn't think of another more specific term. the area is actually some sort of nature reserve, that's another possibility

IMG_0181 IMG_0177

car park

an area that is neither road nor pavement - no traffic flow - but intended to be occupied by parked vehicles . this might be annotatable aswell.. another more specific parking space would do

forest road

IMG_0152

restaurant/cafe .. the outdoor seating for eating outside.. tables, chairs, but in a particular arrangement. very likely people.

IMG_0428 IMG_0378

countryside .. fairly broad catch all term.. maybe road or path, fields, trees, sparse buildings, maybe farm elements

IMG_0293 IMG_0276

docks/port ... the loading/unloading - distinctive gantry cranes, container facilities (shipping container)

IMG_0040

I would also add airport to cover your images (I have one nearby aswell which I might be able to get to).. then perhaps harbour, cityscape, factory,workshop,garage,town centre (less like a street - walkable area, mostly shops), dining room,bedroom,living room,hotel room , residential street (more specific than street), highstreet (street with mostly shops) etc

museum would be a place to get varied images

bbernhard commented 5 years ago

Thanks a lot!

We've talked already a bit here about scene labels. One thing that is not totally clear to me is: When I add a scene label, would the scene label automatically add a bunch of object labels to the image? Or would you see the scene label more like as a way to give hints for labeling and searching? e.g: If the image has the scene label airport we could make the labels plane, bus, aircraft...etc in the labeling screen more prominent. So that those are easier for the user to select.

I can definitely see the latter, but I am not sure about the former. I think automatically adding additional labels based on a scene label could result in quite a few false-positives.

dobkeratops commented 5 years ago

I think I might have posted similar before.. an idea on how to actually use the scene labels in training.. the idea is scene labels give a whole image judgement (for very little annotation work) that should still relate to object labels.. this could be exploited with some sort of shortcut output. During training? Have you seen any other datasets and training methods work like this? IMG_1207

dobkeratops commented 5 years ago

(Tried to summarise a scene label list again.. was just going through trying to add more

bbernhard commented 5 years ago

Thanks a lot for all the suggestions - I've just pushed them to production :)

dobkeratops commented 5 years ago

IMG_2296 Got a tonne of photos from some motor show , some general scene label requests: ‘motorshow’,’festival’,’exhibition’,’motorsport’ .. the whole day was a blend of these things (different areas would correspond to different parts there) Some other useful labels for this image set: “open wheel race car” (“formula car” possibly.. there’s F1,Indycar etc.. open wheel is the general term I think..) ..“super car”, “luxury car”,”racing car”, “car engine”, “racetrack”, “stadium”,”hay bale”,”vintage car”,”car engine” These remind me to ask for the collective object label: “crowd” .. this is gives slightly more specific information than plural “persons” (I.e. density, and a wider space.. eg vs just a small group of people or line of people ina queue)

Got to see the self driving tractor and self driving race car prototypes.. pretty interesting .. plus an awkward example (life size Lego supercar replica , not sure where that sits on the toy vs car debate hah). Makes a change from the usual urban examples

IMG_2280 IMG_2230 IMG_2218 IMG_2191 IMG_2137 IMG_2124