iSEARLE: Improving Textual Inversion for Zero-Shot Composed Image Retrieval

5 May 2024  ยท  Lorenzo Agnolucci, Alberto Baldrati, Marco Bertini, Alberto del Bimbo ยท

Given a query consisting of a reference image and a relative caption, Composed Image Retrieval (CIR) aims to retrieve target images visually similar to the reference one while incorporating the changes specified in the relative caption. The reliance of supervised methods on labor-intensive manually labeled datasets hinders their broad applicability. In this work, we introduce a new task, Zero-Shot CIR (ZS-CIR), that addresses CIR without the need for a labeled training dataset. We propose an approach named iSEARLE (improved zero-Shot composEd imAge Retrieval with textuaL invErsion) that involves mapping the visual information of the reference image into a pseudo-word token in CLIP token embedding space and combining it with the relative caption. To foster research on ZS-CIR, we present an open-domain benchmarking dataset named CIRCO (Composed Image Retrieval on Common Objects in context), the first CIR dataset where each query is labeled with multiple ground truths and a semantic categorization. The experimental results illustrate that iSEARLE obtains state-of-the-art performance on three different CIR datasets -- FashionIQ, CIRR, and the proposed CIRCO -- and two additional evaluation settings, namely domain conversion and object composition. The dataset, the code, and the model are publicly available at https://github.com/miccunifi/SEARLE.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Zero-Shot Composed Image Retrieval (ZS-CIR) CIRCO iSEARLE-XL-OTI (CLIP L/14) mAP@10 12.67 # 8
Zero-Shot Composed Image Retrieval (ZS-CIR) CIRCO iSEARLE (CLIP B/32) mAP@10 11.24 # 10
Zero-Shot Composed Image Retrieval (ZS-CIR) CIRCO iSEARLE-XL (CLIP L/14) mAP@10 13.61 # 4
Zero-Shot Composed Image Retrieval (ZS-CIR) CIRCO iSEARLE-OTI (CLIP B/32) mAP@10 10.94 # 11
Zero-Shot Composed Image Retrieval (ZS-CIR) CIRR iSEARLE-XL (CLIP L/14) R@5 54.00 # 13
Zero-Shot Composed Image Retrieval (ZS-CIR) CIRR iSEARLE-OTI (CLIP B/32) R@5 55.18 # 8
Zero-Shot Composed Image Retrieval (ZS-CIR) CIRR iSEARLE (CLIP B/32) R@5 55.69 # 7
Zero-Shot Composed Image Retrieval (ZS-CIR) CIRR iSEARLE-XL-OTI (CLIP L/14) R@5 54.05 # 12
Zero-Shot Composed Image Retrieval (ZS-CIR) Fashion IQ iSEARLE-OTI (CLIP B/32) (Recall@10+Recall@50)/2 34.93 # 14
Zero-Shot Composed Image Retrieval (ZS-CIR) Fashion IQ iSEARLE-XL (CLIP L/14) (Recall@10+Recall@50)/2 38.24 # 9
Zero-Shot Composed Image Retrieval (ZS-CIR) Fashion IQ iSEARLE-XL-OTI (CLIP L/14) (Recall@10+Recall@50)/2 39.39 # 7
Zero-Shot Composed Image Retrieval (ZS-CIR) Fashion IQ iSEARLE (CLIP B/32) (Recall@10+Recall@50)/2 34.60 # 15
Zero-Shot Composed Image Retrieval (ZS-CIR) ImageNet iSEARLE (CLIP B/32) Average Recall 16.01 # 8
Zero-Shot Composed Image Retrieval (ZS-CIR) ImageNet iSEARLE-XL-OTI (CLIP L/14) Average Recall 22.59 # 2
Zero-Shot Composed Image Retrieval (ZS-CIR) ImageNet iSEARLE-XL (CLIP L/14) Average Recall 24.46 # 1
Zero-Shot Composed Image Retrieval (ZS-CIR) ImageNet iSEARLE-OTI (CLIP B/32) Average Recall 15.62 # 9
Zero-Shot Composed Image Retrieval (ZS-CIR) ImageNet-R iSEARLE-XL (CLIP L/14) (Recall@10+Recall@50)/2 24.46 # 1
Zero-Shot Composed Image Retrieval (ZS-CIR) ImageNet-R iSEARLE-OTI (CLIP B/32) (Recall@10+Recall@50)/2 15.62 # 8
Zero-Shot Composed Image Retrieval (ZS-CIR) ImageNet-R iSEARLE (CLIP B/32) (Recall@10+Recall@50)/2 16.01 # 7
Zero-Shot Composed Image Retrieval (ZS-CIR) MS COCO iSEARLE-OTI (CLIP B/32) Actions Recall@5 26.63 # 6
Zero-Shot Composed Image Retrieval (ZS-CIR) MS COCO iSEARLE (CLIP B/32) Actions Recall@5 26.40 # 7
Zero-Shot Composed Image Retrieval (ZS-CIR) MS COCO iSEARLE-XL-OTI (CLIP L/14) Actions Recall@5 32.55 # 1
Zero-Shot Composed Image Retrieval (ZS-CIR) MS COCO iSEARLE-XL (CLIP L/14) Actions Recall@5 30.05 # 3

Methods