Object tags as alignment anchors

How this idea evolved

Each step is the idea’s introducing paper. The line between steps is the citation link between those papers.

1. 2017
   Bottom-Up Top-Down attentionObject-detector region features

   Represent an image by features of detected object regions instead of a pixel grid.

   No direct link between these papers in this dataset
2. 2020
   OscarObject tags as alignment anchors

   Feed detected object labels as text so they anchor image–text alignment.