Give more meaningful label to the image patches

catherinesyeh / attention-viz

Visualizing query-key interactions in language + vision transformers

http://attentionviz.com/

MIT License

122 stars 15 forks source link

Give more meaningful label to the image patches #56

Closed yc015 closed 1 year ago

yc015 commented 1 year ago

Currently the only annotation we gave to an image patch is its position in the original image, but there are attributes that are way more semantic than that.

[x] Try FCN to get more meaningful label for image patches.

yc015 commented 1 year ago

Done. Currently using FCN automatic label.