• Discover new ways to elevate your game with the updated DGCourseReview app!
    It's entirely free and enhanced with features shaped by user feedback to ensure your best experience on the course. (App Store or Google Play)

A.I. to classify discs from photos - A call for data.

DiscVision

Newbie
Joined
Jan 29, 2017
Messages
3
Hello DGCR, long time disc golfer and aspiring data scientist here.

As most of you are aware, artificial intelligence and deep neural networks have become incredibly powerful at classifying images. The level of accuracy current models from Google, Facebook, etc across the entire spectrum of objects in our world has reached new heights and become incredibly powerful.

What I plan to create:

It is simple, a neural network trained on disc golf discs, to accurately classify the mold of a disc. Simply snap a picture of the profile of a disc, feed it into the network, and have it compute and correctly classify the disc.

I have built a simple classifier to differentiate between Drivers and Putters fairly easily, and with a limited set of images and data I threw together.

What these types of networks need in order to perform is Data. Lots of it. The more the merrier. That is where the disc golf community comes in.

What I need from you:

Pictures of discs, along with the label of what the disc pictured is (i.e Destroyer). At this time, in order to gather as much data as possible, I am choosing to limit the data to ONLY the following popular molds:

Drivers:

- Destroyer
- Wraith
- Teebird
- Trespass
- Felon
- Firebird

Fairway/Midrange:
- Buzzz
- Roc3
- Leopard
- Emac Truth
- River
- Justice

Putters:
- Wizard
- Aviar
- Judge
- Harp
- Dagger
- Nova

-----------------------------------

How to contribute:

Simply take a picture of the discs you want to contribute similar to these examples:

fctwVvq.jpg

mrWKro3.jpg


An even flat side profile of the disc is what I am after, with about half or so of the disc in view. That's it.

A long with the picture, I need to know what the disc is, which should be included in the subject or the body of the email. Or if you want to really help and take profile pictures of your 25 destroyers, you can zip them up or put them in a folder titled destroyer. As long as I have the picture and know what the disc is, I can make use of it and use it to build up the database of pictures for each mold. Just one photo for each individual disc is what I need, please no multiple pictures of the same disc.

If you would like to help out , please email your photographs (from the molds listed above) to:

whatdiscisthis AT gmail DOT com ---- (I am sure you can decode the proper email address)

I am just a guy trying to make a fun cool service to eventually post to a website front-end, and this project will never be monetized, nor will your email ever be used for anything.

Thanks everyone! If anyone has any questions about anything feel free to ask.
 
Appreciate the link but that isn't what I could use.

The idea is to get hundreds (thousands?) of single disc profile pictures for each mold. That is the type of data necessary to train a network to accurately classify molds.
 
Its too bad more of the companies dont do what MVP discs does, because they give us these for every disc.

th
 
I own hundreds of discs and my local disc golf groups and friends were going to contribute as well. I planned to take the pictures of all of my discs of the above molds.

The reason I was asking for the disc golf community to help is because the amount of data required is very large, and the more data the better model.

Having a large number of people take a few minutes out of their day to snap a few disc pictures goes a long way to building a big dataset. In addition the variations in each disc and picture taken goes a long way towards making a better model.

The final model and tool will be released for free and this is the best way to gather data.

This sort of crowdsourcing for data is used across a number of industries and hobbyist groups.
 
Will you need to account for inconsistencies in the ways discs mold up?

I think that's why he wants as many data points as possible. If all Destroyers were the same, you'd only need 1 picture for a program like this. But since there are probably 100s of variations, you need pics of all different kinds to be able to accurately classify any random destroyer.
 
I don't want A.I. thinking about Nukes and Destroyers :rolleyes:
latest
 

Latest posts

Top