[CVPR'26] Air-Know: Arbiter-Calibrated Knowledge-Internalizing Robust Network for Composed Image Retrieval

Zhiheng Fu1, Yupeng Hu1*, Qianyun Yang1, Shiqi Zhang1, Zhiwei Chen1, Zixu Li1,
1Shandong University
*Corresponding author.
Paper Coming Soon Complete source code

Abstract

MY ALT TEXT

NTC Noise Challenges and Our Decoupled Three-Phase Paradigm

MY ALT TEXT

(a) illustrates the semantic ambiguity of noise in NTC. (b) illustrates the vicious cycle of self-dependency caused by unreliable noise determination. (c) introduces our proposed “Expert-Proxy-Diversion” three-phase learning framework. Figure best viewed in color.


Framework: ArbIteR calibrated Knowledge iNternalizing rObust netWork (Air-Know)

MY ALT TEXT

The proposed Air-Know consists of three primary modules: (a) External Prior Arbitration leverages an offline multimodal expert to generate reliable arbitration priors for CIR triplets, bypassing the unreliable small-loss hypothesis. (b) Expert-Knowledge Internalization transfers these priors into a lightweight proxy network, structurally preventing the memorization of ambiguous partial matches. Finally, (c) Dual-Stream Reconciliation dynamically integrates the internalized knowledge to provide robust online feedback, guiding the final representation learning. Figure best viewed in color.

Experiment

MY ALT TEXT

Performance comparison on the FashionIQ validation set in terms of R@K(%). The best and second-best results are highlighted in bold and underline, respectively.


MY ALT TEXT

Performance comparison on the CIRR test set in terms of R@K(%) and Rsub@K(%). The best and second-best results are highlighted in bold and underline, respectively.

MY ALT TEXT

Ablation study on FashionIQ and CIRR datasets. Best and sub-optimal results are highlighted in bold and underline.

MY ALT TEXT

Sensitivity to the hyperparameters (a) p and (b) λ.

MY ALT TEXT

Case Study on (a) CIRR and (b) FashionIQ.

BibTeX


        [Paper BibTex]