Fig 1.
Overview of the proposed method.
Example query from the WikiArt dataset, where to is initial IR results, and tn is Rococo’ conditioned IR results. The proposed method updates embedding vectors to to tn through ti by exploring the embedding space. A green box indicates that the retrieved image has the same class as the condition.
Fig 2.
Proposed conditional image retrieval method.
Fig 3.
Proposed Backward Search algorithm for conditional image retrieval.
Fig 4.
Proposed knowledge distillation model for the Backward Search.
Table 1.
Datasets and queries used in the experiments.
Table 2.
Results of the Backward Search performance without regularization.
Fig 5.
Backward Search results using various λ regularization parameter settings.
(A) t-SNE scatterplot depicting the embedding space of the Baroque and Impressionism dataset embeddings from WikiArt. The ‘init’ query image is marked on the plot. (B) Conditional image retrieval (CIR) result for the query image in (A). The embedding vectors of the top three retrieved images are averaged and visualized on the same t-SNE scatterplot in (B). A green box indicates that the retrieved image has the same class as the condition.
Fig 6.
Variation in performance of mAP@10 and mTS@10 based on the lambda value of the regularization term in the test dataset.
Table 3.
Experiments involving variations in model architectures and backbone fine-tuning.
Table 4.
Comparison study between the proposed method and state-of-the-art CoIR models on the CUB, aPY, and WikiArt datasets.
Fig 7.
Qualitative result of WikiArt test dataset with WikiArt-style query.
We mark correct retrieved images with green boxes for the best view.
Fig 8.
Qualitative result of WikiArt test dataset with WikiArt-multi query.
We mark correct retrieved images with green boxes for the best view.
Table 5.
Iterations and time consumption per query for the Backward Search.
Table 6.
Results of knowledge distillation of the Backward Search.
Fig 9.
Qualitative multiple condition image retrieval results.
We marked the labels of the WikiArt dataset on each retrieved image with colored circles. purple is "sketch and study", green is "landscape", and blue is "Post Impressionism".