A Sparse Feature Matching Model Using a Transformer towards Large-View Indoor Visual Localization

Abstract

Accurate indoor visual localization has been a challenging task under large-view scenes with wide baselines and weak texture images, where it is difficult to accomplish accurate image matching. To address the problem of sparse image features mismatching, we develop a coarse-to-fine feature matching model using a transformer, termed MSFA-T, which assigns the corresponding semantic labels to image features for an incipient coarse matching. To avoid the anomalous scoring of sparse feature interrelationship in the attention assigning phase, we propose a multiscale forward attention mechanism that decomposes the similarity-based features to learn the specificity of sparse features, the influence of position-independence on sparse features is reduced and the performance of the fine image matching in visual localization is effectively improved. We conduct extensive experiments on the challenging datasets; the results show that our model achieves image matching with an average 79.8% probability of the area under the cumulative curve of the corner point error, which outperforms the related state-of-the-art algorithms by an improvement of 13% probability at 1 m accuracy for the image-based visual localization in large view scenes.