Integration of Activation maps into DD about deepdanbooru HOT 4 CLOSED

It already contains command "grad-cam". It estimates all tags in input image and generate gradient map images for each tags. (But lack of post-processing so output image looks not clear)

https://github.com/KichangKim/DeepDanbooru/blob/master/deepdanbooru/commands/grad_cam.py

from deepdanbooru.

Okay, what is the array dimensions of this grad-cam feature? If it is the same as the image itself then I will close this issue, since I expect this to be a weak version of image/semantic segmentation.

from deepdanbooru.

Its dimensions is same to the input of network, 299x299. So if you want to get the value of original image, you need to transform gradient map image to original resolution.

from deepdanbooru.

The major issue is granularity or resolution, because I do have expectations of boundaries being high def.

from deepdanbooru.