Help with my AI Assignment please

SatNavSaysStraightOn

SatNavSaysStraightOn

Moderator
Staff member
Joined
11 Oct 2012
Local time
8:48 PM
Messages
22,485
Location
SE Australia
Website
www.satnavsaysstraighton.com
I need help with one of my assignments please.

I have to train an AI image classifier to recognise and categorise an image according to the following categories.
  1. clear plastic bottle,
  2. opaque plastic bottle,
  3. aluminium beverage can,
  4. steel food can,
  5. glass bottle,
  6. glass jar,
  7. cardboard box,
  8. paper cup.
To do this, I need about 100 photographs with different backgrounds and from different angles of these items for each of these 8 categories (so 800 images in total!). The photographs can only contain items from 1 specific category, but can contain more than 1 item of the same category (so a picture of a supermarket shelf with only tins on it would be fine, but tins and glass jars would not be OK).

The photos can include crushed or partially crushed, opened, unopened, with label, without labels, tins with food and opened, items on their sides, photos of them from underneath etc. It's time to be creative.

With food tins, I'm looking for different shapes and sizes, square, rectangular, pie on a plate, anything that could trip it up.

Paper cups will also be particularly difficult,
And distinguishing aluminium drinks cans from steel food cans may also cause problems, so the more photos the better.

Your images must include meaningful variation. Your dataset should not consist only of highly similar images or repeated standard product photos. Your images should include variation such as: different viewing angles, different lighting conditions, different backgrounds, different object sizes within the image, different object placements or orientations, partial occlusion or non-standard views, and so on.
Click to expand...

And I need them emailed to me please, not posted here because I need to be able to change quality of images. Though if email is out of the question but you'd still like to participate, we can use the media gallery to transfer them but the MG will limit the size of each image to 400Kb, hence why email would be preferred if possible.

Getting images from around the world will also help ensure that my training dataset is not Australian biased.

If you are interested in helping me, please PM me for my university email address. I will need those helping out to sign a form allowing me to use their images to train the AI (Google Teachable Machine) for the assignment. The images will not be used beyond that, and I will need to included the release form in the appendix of my assignment.

If I use photos from family and friends, I'll avoid copyright issues, plus everyone photographs things differently so I'll get a good variety of images, from a variety of different devices and qualities. If you don't want me knowing your name, I'm happy for the "signature" on the release form to be tied to your CB username.

Even a broken glass bottle or discarded paper cup at the side of the road is fine!

All images will be stripped of metadata before use and if there is anything that may identify you (pill container, wedding ring), that will be removed from the image before use.

Any one interested please?

The time frame is maximum 14 days to create the dataset of 800 images. You can send just 1 or 2 or +50 images, i honestly don't mind.
 
badjak said:
Are pictures taken with a (cheapish) cell phone good enough?
Click to expand...
Yes, absolutely.

It is just uploading them here, resizes them, whereas with email, that can be overridden and they will be larger in size which is preferable. Here we end up with 200Kb in a post, 400Kb in the media gallery, but by email, I'll receive the size you took the image at, if you allow email to send the original size, not resize then before sending.
 
karadekoolaid said:
Do the items need to be empty, SatNavSaysStraightOn, or can they contain something?
If I take a pic of some tins/jars/bottles in the cupboard, for example, will they do?
Click to expand...
any which way. so full unopened, partially used, empty not washed, empty washed...
The only restriction I have found out about today, whilst trying to add images to the training app, is that they need to be in a 1:1 format which you can usually kick your phone into quite easily - if it says 4:3, just tap on the 4:3 and you should be given some extra options - whatever your lowest quality is is fine as well, doesn't need to be 50Megapixels or anything of fine art print quality.

Ideally, for training, the item should be by itself, or with like for like - so all glass jars. I can't have jars and bottles mixed together.

I just pulled a few things out of the fridge and put them on the counter to take some pictures.
 
You must log in or register to reply here.
Back
Top Bottom