DeAR: Fine-Grained VLM Adaptation by Decomposing Attention Head Roles

Benchmark Model Rank Results
prompt-engineering-on-dtdDeAR#1Harmonic mean: 79.62
prompt-engineering-on-eurosatDeAR#1Harmonic mean: 92.22
prompt-engineering-on-sun397DeAR#1Harmonic mean: 82.80
prompt-engineering-on-imagenet-aDeAR#2Top-1 accuracy %: 51.80
prompt-engineering-on-imagenet-rDeAR#2Top-1 accuracy %: 78.83
prompt-engineering-on-fgvc-aircraftDeAR#3Harmonic mean: 42.61
prompt-engineering-on-oxford-102-flowerDeAR#3Harmonic mean: 87.57
prompt-engineering-on-imagenetDeAR#4Harmonic mean: 74.83
prompt-engineering-on-stanford-carsDeAR#4Harmonic mean: 78.47
prompt-engineering-on-caltech-101DeAR#5Harmonic mean: 96.95
prompt-engineering-on-imagenet-v2DeAR#6Top-1 accuracy %: 64.87
prompt-engineering-on-oxford-iiit-pet-datasetDeAR#8Harmonic mean: 96.81
prompt-engineering-on-imagenet-sDeAR#10Top-1 accuracy %: 49.50