Original Post
DINOv2 SALAD is the SOTA single-stage VPR (Extractor-Retriever-Matcher for images). It uses Optimal Transport DUSTBIN, just as SuperGlue did in the late 2019.
What if we go away with DUSTBIN, just as it was done in LightGlue? Can the Transformer be improved even further? Philipp Lindenberger | Paul-Edouard Sarlin | Marc Pollefeys | Javier Civera | Tomasz Malisiewicz | Daniel DeTone | Andrew Rabinovich
🤔
Links:
* DIVOv2 SALAD (the cover): https://lnkd.in/gvgwd4dz
* SuperGlue: https://lnkd.in/giTDQBtA
* LightGlue: https://lnkd.in/gzsjuvyq