Better and Faster: Exponential Loss for Image Patch Matching

VEDesc: vertex-edge constraint on local learned descriptors

“…L2Net [13] designs a deeper network and produces descriptor normalized to the unit norm by L 2 distance. The architecture of L2Net is an application and foundation of learned descriptors in the later works [14,16,17]. However, the metric-learning loss function of L2Net is less effective in finding hard samples from negative and anchor samples.…”

“…Many previous works of metric learning focused on learning Mahalanobis distance [31,32], while much more efforts are spent on the learning vectors with Euclidean distance metrics [33,34]. Particularly, several main studies [14,16,17] , which adopt L2Net structure, redefine the improved triplet loss function with Euclidean distance metric to improve performance of learned descriptors. HardNet [14] introduces a new triplet loss into L2Net, which can minimize the distance between the matching descriptors and closest non-matching descriptors.…”

“…DOAP [15] proposes optimizing average precision (OAP) to L2Net and achieves more competitive results on many datasets. ExpTLoss [16] introduces the exponential Siamese and triplet loss function into L2Net, and achieve better performance on many tasks. CDF [17] assigns a nonparametric soft margin instead of a hard margin in L2Net and improves the performance of learned descriptors.…”

“…However, handcrafted descriptors are limited in terms of robustness and precision due to the lack of high-level structural information. Recently, with the development of convolutional neural networks (CNNs), the research hotspot focuses on the learned descriptors [13][14][15][16][17] instead of the handcrafted descriptors. These learned descriptors achieve much higher performance than the handcrafted ones in various areas, including image matching and patch verification.…”

“…Specifically, To improve the performance of learned descriptors, the triplet loss [14] encourages that the distance between matching descriptors is less than the distance between nonmatching ones. Inspired by the triplet loss, recent works [16,17] try to minimize the distance between matching descriptors and maximize the distance between non-matching descriptors. These experimental results demonstrate the triplet loss is effective on various datasets, improving the robustness of learned descriptors in extreme conditions, including weather, season, illumination, and distortion.…”