-
Notifications
You must be signed in to change notification settings - Fork 1
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Interest detection similar to Detectron2 #3
Comments
Another concept: Recaptcha-like image regional refinement, for example present a 3x3 square of images for people to tag, where 5 of the images are not related to the tag in question, 3 of the images are human-affirmed images with the tag, and 1 single image where it is machine tagged to be true but not yet confirmed by a human (5-3-1 strategy). In order to prevent bounding box bias in image orientation (+-45degrees) the images would have to be altered slightly such that people can tag them without the influence of background noise. Also in order to prevent overfitting and ML botting, "Adversarial Examples" would need to be implemented in order to fuzz things and make training better (in a reasonable manner that is not too extreme when compared distance in YCbCr or similar color system). |
It really isn't made for segmentation, so I'm not sure how one could effectively get super nice proper segment borders. Wrt resistance to adversarial examples: This is a published model, so you're in a white-box situation - and as far as my understanding goes, that means you're fairly close to hosed no matter what - in which case it makes more sense to specifically look at where you'd want to use the model, what your expected threats are, and how to mitigate them other than by trying to make the model a tiny bit better (since you'll want those semi-automatic, non-ml mitigations anyways). In any case, I see you did find the original authors repo, so I'll close this here - but feel free to fork off or copy my code if it's useful for your segmentation experiments. |
In detectron2 (see ) you can breakdown an image based on what objects or parts or items are inside an image... is it possible to apply this to DDAM through image edge detection?
The text was updated successfully, but these errors were encountered: