Learned descriptor using dynamical exponential algorithm

Inhalt:
Recent works improve the performance of learned descriptors by the triplet loss function and make some effort in finding the hard negative samples. However, these methods pay less attention to the weight parameter of the loss function to improve the stability of learned descriptors. Unstable descriptors can cause mismatching. To solve the problem, we design a dynamical exponential algorithm (DEA) and introduce the algorithm into the triplet loss function. This algorithm can make that the distance between nonmatching descriptors is exceeded to the distance between matching descriptors. It can improve the stability of the learned descriptor by increasing the difference between two distances. We demonstrate that the performance of the consistent descriptors reached state-of-the-art on various datasets.