CoLLM: A Large Language Model for Composed Image Retrieval

Benchmark Model Rank Results
zero-shot-composed-image-retrieval-zs-cir-on-circoCoLLM (Pretrained - CLIP-L/14)#20mAP@10: 20.8MAP@5: 20.3mAP@50: 23.4
zero-shot-composed-image-retrieval-zs-cir-on-circoCoLLM (Pretrained - BLIP-L/16)#21mAP@10: 20.4MAP@5: 19.7mAP@50: 23.1
zero-shot-composed-image-retrieval-zs-cir-on-cirrCoLLM (finetuned - BLIP-L/16)#1R@1: 45.8R@10: 84.7R@50: 95.8
zero-shot-composed-image-retrieval-zs-cir-on-cirrCoLLM (Pretrained - BLIP-L/16)#7R@1: 35.00R@10: 78.6R@50: 94.2
zero-shot-composed-image-retrieval-zs-cir-on-cirrCoLLM (Pretrained - CLIP-L/14)#13R@1: 29.7R@10: 72.8R@50: 91.5
zero-shot-composed-image-retrieval-zs-cir-on-fashion-iqCoLLM (finetuned - BLIP-L/16)#4(Recall@10+Recall@50)/2: 49.9R@10: 39.1R@50: 60.7
zero-shot-composed-image-retrieval-zs-cir-on-fashion-iqCoLLM (Pretrained - BLIP-L/16)#12(Recall@10+Recall@50)/2: 45.3R@10: 34.6R@50: 56.0
zero-shot-composed-image-retrieval-zs-cir-on-fashion-iqCoLLM (Pretrained - CLIP-L/14)#24(Recall@10+Recall@50)/2: 39.8R@10: 30.1R@50: 49.5