[TIP'26] COMBINER: Composed Image Retrieval Guided by Attribute-based Neighbor Relations

Zixu Li¹, Yupeng Hu^1*, Zhiwei Chen¹, Haokun Wen^2,3, Xuemeng Song⁴, Liqiang Nie²

¹School of Software, Shandong University,
²School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen),
³School of Data Science, City University of Hong Kong,
⁴Department of Computer Science and Engineering, Southern University of Science and Technology
^*Corresponding author.

Paper Coming Soon Complete source code

Illustration of Relations Modeling

Example of (a) Pairwise Relations, (b) Neighbor Relations, and (c) Visually Similar Images in Relations Modeling. In this figure, Q denotes the multimodal query, T denotes the target image, and C denotes the candidate image. Fig. 1(c) illustrates the traditional neighbor relations modeling methodology brings both candidate images C₁ and C₂ close to Q₁. However, C₂ is visually similar but attribute-unrelated with Q₁ ("carpet" does not match the query "bedding"). Therefore, C₂ should not be brought close to Q₁.

Attribute Prototype-Based Similarity Measure

Schematic of our proposed similarity measure method based on attribute prototypes

Framework: COMposed image retrieval network guided By attrIbute-based NEighbor Relations (COMBINER)

Overall framework of COMBINER, which consists of (a) Adaptive Semantic Disentanglement, (b) Unified Prototype-based Composition, and (c) Dual Relations Modeling.

Experiment

Performance comparison on the FashionIQ dataset with respect to R@k (%). The top-performing results across all methods are highlighted in blue, and the best results among baseline methods are indicated with underlining.

Performance comparison on CIRR with respect to R@k(%) and Rsubset@k(%). The overall best results are colored in blue, while the best results over baselines are underlined.

Performance comparison on Shoes with respect to R@k(%). The overall best results are colored in blue, while the best results over baselines are underlined.

Ablation Studies of COMBINER with different components and various settings on FashionIQ, Shoes, and CIRR.

Influence of hyper-parameters ρ, κ, and μ on FashionIQ and CIRR datasets.

Influence of (a) Attribute Prototype Number U and (b) Semantic Cluster Number H on FashionIQ and CIRR datasets.

Case study on (a) FashionIQ, (b) Shoes, (c) CIRR, and failure cases on (d) CIRR and (e) FashionIQ.

Attention Visualizations on (a) Dresses, (b) Shirts, (c) Tops&Tees, and (d) CIRR datasets.

Visualization of Semantic Cluster Neighbors on (a) FashionIQ and (b) CIRR datasets.