Image recognition of racehorses

[brogers]

I'm working on an image recognition program that can classify yearling photos using a convoluted neural network. Here is the first go at it for all to take a look at.
[url]
https://www.metamind.io/classifiers/40114[/url]

All you have to do is either upload an image of a yearling in a standard pose or a link of the same.

It is a binary outcome so while it gives a percentage chance of being either elite or non-elite it will select one over the other. It isnt that well tuned as its accuracy is just a little over 70% but that is a result of sample size. You'd really like it to train on at least 1000 images which I will get to. I also need to add more images from North America as most are off Australian/NZ sales company sites. If anyone is interested here is the background of CNN's https://en.wikipedia.org/wiki/Convolutional_neural_network

Feedback most appreciated.

[Tappiano]

It likes my yearling and his informal pose.
Elite Racehorse 53%
Non Elite Racehorse 46%

[brogers]

The model increases accuracy if you select the button next to the class (elite or not elite) that the horse actually is.

[Tappiano]

I uploaded three pictures of him. The first one gives the same % mentioned above, even when I check non-elite. The second picture is him but he's standing ever so slightly different, head held a tad higher, not to mention he's got his mouth partially open and then it swings to 72% non and 27 elite. The third one which from the other side is 64 E, 35 N with the N checked. I then uploaded one of my mare and it was 65/35 e vs n so while the mare and the offspring are attractive they aren't elite.

[brogers]

It suffers a little through lack of sample size. The 500 or so just isnt enough but I am working on a larger and more accurate model that uses close to 10,000 images to learn on. That should be more accurate and more adaptable to horses not standing in a traditional pose.

[Joltman]

cool stuff Byron

[cewright]

sent a pm awaiting response