DeepStack model building >>> Is there a way to include "negative" confirmation images

[Sybertiger]

If you build DS models then you know you include data to train and test the model. And, the images you use are to positively identify objects. I think we've all seen cats identified as dogs or a person. Is there a way to include images to train DS to tell it that a cat isn't a dog or a person? In other words, submit training data of a dog or person and let DS know "this isn't a cat"?

 

[VideoDad]

If you build DS models then you know you include data to train and test the model. And, the images you use are to positively identify objects. I think we've all seen cats identified as dogs or a person. Is there a way to include images to train DS to tell it that a cat isn't a dog or a person? In other words, submit training data of a dog or person and let DS know "this isn't a cat"?

Yes, just include the image in the training set but without the tag causing the false positive.

 

[Sybertiger]

So, you would include the false positive image and give it a tag of what it actually is. For example, if you were building a dog specific model you would include the pic of the false positive of the cat (DS identified as a dog) and tag it as "cat"?

 

[VideoDad]

So, you would include the false positive image and give it a tag of what it actually is. For example, if you were building a dog specific model you would include the pic of the false positive of the cat (DS identified as a dog) and tag it as "cat"?

If you have a model that only identifies dogs, then you would include pictures of cats, skunks, squirrels, etc. and don't tag them at all.

If you have a model of cats and dogs and it continually gets a certain set of cats and thinks they are dogs, you'd add as many of those images tagged correctly as cat.

I had to do this with my delivery model that was seeing certain shrubs, car wheels, etc. as the wrong thing. I included those images that had previously been misidentified and included them with no tag so it learns they aren't the desired object.