MagicLens: Self-Supervised Image Retrieval with Open-Ended Instructions

Benchmark Model Rank Results
zero-shot-composed-image-retrieval-zs-cir-on-circoMagicLens (CoCa L)#4mAP@10: 35.4
zero-shot-composed-image-retrieval-zs-cir-on-circoMagicLens (CoCa B)#8mAP@10: 32.0
zero-shot-composed-image-retrieval-zs-cir-on-circoMagicLens (CLIP L)#10mAP@10: 30.8
zero-shot-composed-image-retrieval-zs-cir-on-circoMagicLens (CLIP B)#17mAP@10: 23.8
zero-shot-composed-image-retrieval-zs-cir-on-cirrMagicLens (CoCa L)#9R@1: 33.3R@5: 67.0R@10: 77.9R@50: 94.4
zero-shot-composed-image-retrieval-zs-cir-on-cirrMagicLens (CLIP L)#12R@1: 30.1R@5: 61.7R@10: 74.4R@50: 92.6
zero-shot-composed-image-retrieval-zs-cir-on-cirrMagicLens (CoCa B)#24R@5: 64.0
zero-shot-composed-image-retrieval-zs-cir-on-cirrMagicLens (CLIP B)#26R@5: 58.0
zero-shot-composed-image-retrieval-zs-cir-on-fashion-iqMagicLens (CoCa L)#6(Recall@10+Recall@50)/2: 48.1R@10: 38R@50: 58.2
zero-shot-composed-image-retrieval-zs-cir-on-fashion-iqMagicLens (CoCa B)#11(Recall@10+Recall@50)/2: 45.3
zero-shot-composed-image-retrieval-zs-cir-on-fashion-iqMagicLens (CLIP L)#20(Recall@10+Recall@50)/2: 41.6R@10: 30.7R@50: 52.5
zero-shot-composed-image-retrieval-zs-cir-on-fashion-iqMagicLens (CLIP B)#31(Recall@10+Recall@50)/2: 36.85
zero-shot-composed-image-retrieval-zs-cir-on-imagenet-rMagicLens (CoCa L)#1(Recall@10+Recall@50)/2: 52.95
zero-shot-composed-image-retrieval-zs-cir-on-imagenet-rMagicLens (CoCa B)#2(Recall@10+Recall@50)/2: 51.75
zero-shot-composed-image-retrieval-zs-cir-on-imagenet-rMagicLens (CLIP L)#3(Recall@10+Recall@50)/2: 48
zero-shot-composed-image-retrieval-zs-cir-on-imagenet-rMagicLens (CLIP B)#4(Recall@10+Recall@50)/2: 36.4