"LightGlue: Local Feature Matching at Light Speed" ICCV, 2023 Jun paper code pdf Authors: Philipp Lindenberger, Paul-Edouard Sarlin, Marc Pollefeys
- Task
- Background
- 🏷️ Label:
the inference is much faster on image pairs that are intuitively easy to match
adaptive stopping mechanism gives a fine-grained control over the speed vs. accuracy trade-off
设置多个 layer 去对特征点筛选,每层内部用分类器 C 判断是否继续匹配,有效避免不相似的块继续浪费资源匹配。
n find out early that some points (?) are unmatchable and thus exclude them from subsequent lay
同一视频片段内部,差异较小,特征点可以复用,只需学习微小偏移
ablation study 看那个模块有效,总结一下
- early stopping
learn what & how to apply to our task