Spair71k Results

We show two randomly chosen keypoint correspondences from each category of Spair71k subset (see Tab. 4). Our method produces more coherent matches than NBB [1], and approaches the performance of CATs [2] which has been trained in a supervised manner on annotated corresponding pairs.   

aeroplane

Source NBB Ours CATs Ground Truth

cat

Source NBB Ours CATs Ground Truth

bird

Source NBB Ours CATs Ground Truth

motorbike

Source NBB Ours CATs Ground Truth

dog

Source NBB Ours CATs Ground Truth

sheep

Source NBB Ours CATs Ground Truth

boat

Source NBB Ours CATs Ground Truth

bus

Source NBB Ours CATs Ground Truth

car

Source NBB Ours CATs Ground Truth

horse

Source NBB Ours CATs Ground Truth

chair

Source NBB Ours CATs Ground Truth

cow

Source NBB Ours CATs Ground Truth

person

Source NBB Ours CATs Ground Truth

pottedplant

Source NBB Ours CATs Ground Truth

train

Source NBB Ours CATs Ground Truth

tvmonitor

Source NBB Ours CATs Ground Truth

bicycle

Source NBB Ours CATs Ground Truth

bottle

Source NBB Ours CATs Ground Truth

References

[1] Neural Best Buddies: Sparse Cross-Domain Correspondences. Aberman et al. TOG 2018

[2] Cost Aggregation Transformers for Visual Correspondences. Cho et al. NeurIPS 2021