Original post
Dmytro Mishkin 🇺🇦#1505
Dmytro Mishkin 🇺🇦#1505
Johan Edstedt@Parskatt
Correspondence, correspondence, correspondence!
1:51 PM · Jun 4, 2026 · 1.1K Views
Talk Argues Correspondence Drives Representation Learning More Than ImageNet Accuracy
--0--
Sentiment
Sentiment building, check back later.
Cluster Engagement
-
Views
-
Comments
-
Reposts
-
Bookmarks
Expand data
Posts from X
Most Activity
Most Activity
RETWEETS1
Adam Kortylewski 🚨 Hiring PhD students@AdamKortylewski
And it's not just a curiosity.
Across 37 vision backbones, correspondence quality correlates more strongly with downstream tasks — segmentation, 3D pose, tracking, depth — than ImageNet kNN does.
The capability driving everything downstream is the one we measure least carefully
1dViews 733Likes 5Bookmarks 0
