HPatches
A standard benchmark dataset in computer vision used to evaluate local feature matching and same-instance point correspondence.
Timeline (last 90 days)
Redistributing the cross-image attention of 16 to 32 heads (keeping the total unchanged) causes synthetic-task accuracy to drop by 17 to 22 points and HPatches by 15 to 32 points, whereas random heads or attention-matched heads cause at most a 7-point drop.
Training only the key heads' query and key projections outperforms layer-matched-head and attention-matched-head control groups and transfers to HPatches and BLINK.
All Facts (2)
Redistributing the cross-image attention of 16 to 32 heads (keeping the total unchanged) causes synthetic-task accuracy to drop by 17 to 22 points and HPatches by 15 to 32 points, whereas random heads or attention-matched heads cause at most a 7-point drop.
50%UnverifiedTraining only the key heads' query and key projections outperforms layer-matched-head and attention-matched-head control groups and transfers to HPatches and BLINK.
50%