Backward induction-based deep image search.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Conditional Image Retrieval (CIR) is crucial for efficient image search and analysis.
• Existing methods like Composed Image Retrieval (CoIR) require expensive datasets (triplet or image-text pairs).

Purpose of the Study:

• To develop a novel CIR method that bypasses the need for extensive datasets.
• To enable CIR at the image-level concept using an inverse mapping approach.

Main Methods:

• Proposed Backward Search method updates query embeddings to match specified conditions.
• Employs an inverse mapping approach to leverage the model's inductive knowledge.
• Utilizes knowledge distillation to significantly reduce computation time.

Main Results:

• Backward Search achieves an average mAP@10 of 0.541 on WikiArt, aPY, and CUB datasets, surpassing CoIR methods.
• Knowledge distillation enables student models up to 160x faster with minimal performance loss.
• Demonstrates effective single and multi-conditional image retrieval.

Conclusions:

• Backward Search offers an efficient and effective solution for conditional image retrieval using image-level concepts.
• The method reduces reliance on costly annotated datasets.
• Knowledge distillation provides a practical pathway for deploying faster CIR models.