Memory-Based Jitter: Improving Visual Recognition on Long-Tailed Data with Diversity in Memory

This paper considers deep visual recognition on long-tailed data. To make our method general, we tackle two applied scenarios, i.e. , classification and metric learning. Under the data distribution, most classes (i.e., tail classes) only occupy relatively few samples are prone to lack of within-class diversity. A radical solution is augment with higher this end, introduce a simple reliable named Memory-based Jitter (MBJ). We observe that during training, model constantly changes its parameters after every iteration, yielding phenomenon weight jitters. Consequentially, given same image as input, historical editions generate different features in deeply-embedded space, resulting feature Using memory bank, collect these (model or feature) jitters across multiple training iterations get so-called Jitter. The accumulated enhance diversity for consequentially improves recognition. With slight modifications, MBJ applicable fundamental tasks, i.e., learning (on data). Extensive experiments five benchmarks demonstrate significant improvement. Moreover, achieved performance par state art both tasks.