Learning rebalanced human parsing model from imbalanced datasets

摘要

Research on human parsing methods has attracted increasing attention in a wide range of applications. However, dataset imbalance is still a challenging problem in this task, which directly affects the performance of human parsing. There are different types of dataset imbalance problems. For example, the numbers of samples for various labels in a dataset may differ, the scales of objects identified by different labels may vary considerably, the differences between some heterogeneous label types may be much smaller than other cases, and in some extreme situations, images may be labeled incorrectly. In this paper, we propose a rebalanced model for imbalanced human parsing. Two innovative blocks are included in the model, i.e., a pre-bilateral awareness block and a combined-order statistics awareness block. The function of the former is to leverage the multiscale feature extractors to capture the changing scale information in an efficient way from the spatial space. Meanwhile, the function of the latter is to exploit the information of the feature distributions from the channel space. Furthermore, we propose an imbalance data-drop algorithm to simultaneously solve the mislabeling and small sample label weighting problems. Extensive experiments are conducted on three datasets, and the experimental results demonstrate that our method is able to solve the problem of data imbalance efficiently and obtain better human parsing performance.