TopicFM: Robust and Interpretable Topic-Assisted Feature Matching

This study addresses an image-matching problem in challenging cases, such as large scene variations or textureless scenes. To gain robustness to situations, most previous studies have attempted encode the global contexts of a via graph neural networks transformers. However, these do not explicitly represent high-level contextual information, structural shapes semantic instances; therefore, encoded features are still sufficiently discriminative We propose novel method that applies topic-modeling strategy images. The proposed trains latent instances called topics. It models image multinomial distribution topics, and then performs probabilistic feature matching. approach improves matching by focusing on same areas between In addition, inferred topics provide interpretability for results, making our explainable. Extensive experiments outdoor indoor datasets show outperforms other state-of-the-art methods, particularly cases.