Reproduce “Robust Remote Sensing Image-Text Retrieval with Noisy Correspondence” on OpenResearch

About this paper

Qiya Song, Yiqiang Xie, Yuan Sun, Renwei Dian, Xudong Kang

A robust paradigm for remote sensing image-text retrieval (RRSITR) was introduced to mitigate noisy correspondences in training datasets. This method consistently outperforms state-of-the-art techniques, demonstrating enhanced stability and achieving a mean Rank of 35.93% on RSITMD even with an 80% noise rate.

Key insights

• Adopting a human-like cognitive learning progression, from easy to hard examples, through dynamic sample categorization and adaptive weighting effectively filters noise and guides robust cross-modal learning.
• Dynamically adjusted margins within triplet loss functions are crucial for preventing models from overfitting to mislabeled positive examples, especially when dealing with high levels of noise.
• The combination of global and local feature alignment with noise-robust learning objectives, such as self-paced learning and robust triplet loss, enables models to achieve superior and stable performance in noisy remote sensing datasets.

Summary by alphaXiv.