DNN Flow: DNN Feature Pyramid based Image Matching
In Proceedings British Machine Vision Conference 2014
AbstractWe propose a hierarchical image matching algorithm based on DNN feature pyramid, named as DNN Flow. In DNN feature pyramid, the activations of different level response different information, such as semantic on top level and low-level visual appearance on bottom level. DNN Flow utilize this characteristic of DNN feature to achieve different matching goals on different levels. From the top matching level to bottom matching level, DNN flow match the regions with same semantic to the patches with same low-level visual appearance. Meanwhile, the estimation produced by upper level matching lead lower level matching to make more accurate matching. DNN Flow shows competitive evaluation result through various matching tasks on different dataset, which is appropriate to be applied many vision applications.
FilesExtended Abstract (PDF, 1 page, 128K)
Paper (PDF, 10 pages, 1.3M)