LightGlue
This project introduces a high-speed and precise local feature matcher using LightGlue, a deep neural network optimized for matching sparse features across image pairs. Employing adaptive pruning for network width and depth, it delivers fast inference for simple image pairs while effectively managing complexity for more challenging ones. The project includes pretrained weights compatible with SuperPoint, DISK, ALIKED, and SIFT features. The repository provides all necessary code for training and evaluation, delivering faster performance, particularly with the integration of FlashAttention and PyTorch compilation. Experience high-performance feature matching in visual localization and more.
