EVENT DOSSIER
The researchers proposed a method called MULVec, aimed at solving the problem of zero-sample combined image retrieval in an image library without training. This method generates structured query records containing four roles: global, expected to be retained, to be kept, and prohibited. By using a frozen encoder, queries are transformed into description vectors and role probe vectors. The system performs retrieval tasks based on shared evidence and sorts the entire image library with fixed weighted sum scoring. Experiments show that on the CIRCO, CIRR, and FashionIQ datasets, MULVec increased CIRCO mAP@5 by up to 23.0% compared to the strongest competing methods, and achieved the best results in the CIRR and FashionIQ tests.
提出 MULVec 方法，实现无需训练即可在图像库中根据参考图像和文本编辑检索目标图像。该方法通过编译器生成结构化查询记录，映射至全局、期望保留、保持及禁止四个检索角色，利用冻结编码器将查询转化为描述向量与角色探针向量，同时候选图像由全局视觉向量和局部视觉向量库表征。检索角色基于共享证据分别执行任务，最终通过固定加权求和评分对整张图库进行单次检索排序。在 CIRCO、CIRR 及 FashionIQ 数据集及三种骨干网络规模上，MULVec 相比最强对比方法将 CIRCO mAP@5 提升最高 23.0%，并在 CIRR 和 FashionIQ 测试中取得最佳结果。