[CVPR'26] ConeSep: Cone-based Robust Noise-Unlearning Compositional Network for Composed Image Retrieval

Zixu Li1, Yupeng Hu1*, Zhiwei Chen1, Mingyu Zhang1, Zhiheng Fu1, Liqiang Nie2,
1Shandong University,
2Harbin Institute of Technology (Shenzhen)
*Corresponding author.
Paper Coming Soon Complete source code

Abstract

MY ALT TEXT

Sample Types and CIR Challenges in the NTC Scenario

MY ALT TEXT

(a) illustrates examples of “Clean Sample”, “Partial Match Sample” and “Hard Noisy Sample” within the NTC scenario. (b) illustrates the three challenges that current CIR methods face in the NTC scenario, including (b1) Modality Suppression, (b2) Negative Anchor Deficiency, and (b3) Unlearning Backlash.


Framework: Cone-based robuSt noisEunlearning comPositional network (ConeSep)

MY ALT TEXT

The proposed ConeSep consists of three primary modules: (a) Geometric Fidelity Quantization, (b) Negative Boundary Learning, and (c) Boundary-based Targeted Unlearning. In (a), we visualize the similarity cone effect, where the x and y axes denote the similarity scores of clean samples and NTC samples, respectively, and the z axis shows the frequency distribution of these similarity values. The light green region in the figure forms a distinctly cone-shaped space.

Experiment

MY ALT TEXT

Performance comparison on the FashionIQ validation set in terms of R@K(%). The best and sub-optimal 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.

MY ALT TEXT

Sensitivity to (a) Fidelity threshold ω and (b) κ of ℒul .

MY ALT TEXT

Ablation study on FashionIQ and CIRR datasets.

BibTeX


        [Paper BibTex]