Causal Inference Hashing for Long-Tailed Image Retrieval.

Area of Science:

• Computer Vision and Machine Learning
• Information Retrieval

Background:

• Long-tailed bias in image retrieval hinders learning for data-poor classes.
• Existing methods fail to fully leverage or address the dual nature of this bias.
• Causal inference offers a novel perspective to disentangle bias effects.

Purpose of the Study:

• To propose a novel hashing framework using causal inference for long-tailed image retrieval.
• To disentangle detrimental bias effects from beneficial prior knowledge in long-tailed datasets.
• To enhance the learning of discriminative hash codes for both head and tail classes.

Main Methods:

• Developed a hash framework employing causal inference to separate bias effects.
• Constructed hash mediators to capture beneficial prior knowledge from class centers.
• Introduced a de-biased hash loss function utilizing hash mediators and backdoor adjustment.

Main Results:

• The proposed method effectively disentangles beneficial bias from detrimental bias.
• Hash mediators successfully preserve valuable prior knowledge from class centers.
• The de-biased hash loss enhances discriminative hash code learning.
• Significant improvements in retrieval performance demonstrated across four datasets.
• Outperformed state-of-the-art methods by substantial margins.

Conclusions:

• Causal inference provides a powerful approach to address long-tailed bias in image retrieval.
• The proposed framework effectively enhances hash code learning for imbalanced datasets.
• This method offers a promising direction for future research in long-tailed retrieval.