Could I use captum for object localisation? about captum HOT 2 OPEN

Thank you @batrlatom ! I understand! In case of attribution we do not necessarily need to attribute to the labeled classes , we can also attribute to the predicted ones. Do you have any examples and links that you think will be most useful, we will find time and prioritize it ?
Thank you :)

from captum.

Hi,
this is great that you find such usecase usefull! The tutorial could be as easy as making gradcam from input image and thresholding inputs so I can obtain most likely bounding box for the predicted 1-n classes to more advanced like using captum in conjuction with attention dropout layer ( https://github.com/junsukchoe/ADL ). I personally would be most interested if and how could I use captum for learning how to train image cropping in the visual search application .. https://alitech-private.oss-cn-beijing.aliyuncs.com/1534297833520/VisualSearch.pdf see Section 3.2 .

from captum.