A dual dynamic threshold tuning strategy for deep metric learning
Authors: Xiruo Jiang, Yazhou Yao, Sheng Liu, Fumin Shen, Liqiang Nie, Xiansheng Hua
Abstract: Loss functions and sample mining strategies are key components of deep metric learning algorithms. However, existing loss functions or mining strategies often require incorporating additional hyperparameters, especially thresholds that define whether a sample pair is informative or not. The threshold provides a stable numerical criterion for deciding whether to keep a pair or not. It is a key parameter for reducing redundant sample pairs participating in training. However, finding the optimal threshold is a time-consuming task and often requires extensive grid searches. As the threshold cannot be dynamically adjusted during the training phase, determining the threshold requires many repeated experiments. Therefore, we introduce a novel approach to adjust the thresholds related to both the loss function and the sample mining strategy. We design a static Asymmetric Sample Mining Strategy (ASMS) tailored to the sample mining method and its dynamic version, Adaptive Tolerance ASMS (AT-ASMS). ASMS uses differentiated thresholds to address the problems (too few positive pairs, too many negative pairs) that arise from applying only a single threshold for filtering samples. AT-ASMS can adaptively adjust the ratio of positive and negative pairs during training depending on the ratio of positive and negative pairs currently being mined. This meta-learning based threshold generation algorithm uses a single-step gradient descent method to obtain a new threshold. These two threshold adjustment algorithms are combined to form the Dual Dynamic Threshold Adjustment Strategy (DDTAS). Experimental results show that this algorithm achieves competitive performance on CUB200, Cars196, and SOP datasets.
