Less is More: Concise Retrieval and Critic-based Re-ranking for Composed Image Retrieval

Key points

• Recent approaches commonly rely on large vision-language models to generate target image captions for retrieval.1
• However, these captions may include redundant semantic details, introducing noise when retrieving the target image and leading to suboptimal retrieval results.1
• To address this issue, we propose Less-is-More CIR, a two-stage training-free framework that combines Concise Retrieval and Critic Re-ranking.1
• In the first stage, concise target-oriented captions are generated that preserve only the most discriminative modification semantics, thereby reducing semantic noise in the retrieval.1

Sources (1)

• [1]Less is More: Concise Retrieval and Critic-based Re-ranking for Composed Image Retrieval
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 02:53 AM

  Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image based on a reference image and modification text.

  Recent approaches commonly rely on large vision-language models to generate target image captions for retrieval.

Extractive summary: sentences quoted from the sources.