Abstract

Pedestrian attribute information is important function for an advanced driver assistance system (ADAS). Pedestrian attributes such as body pose, face orientation and open umbrella indicate the intended action or state of the pedestrian. Generally, this information is recognized using independent classifiers for each task. Performing all of these separate tasks is too time-consuming at the testing stage. In addition, the processing time increases with increasing number of tasks. To address this problem, multitask learning or heterogeneous learning is performed to train a single classifier to perform multiple tasks. In particular, heterogeneous learning is able to simultaneously train a classifier to perform regression and recognition tasks, which reduces both training and testing time. However, heterogeneous learning tends to result in a lower accuracy rate for classes with few training samples. In this paper, we propose a method to improve the performance of heterogeneous learning for such classes. We introduce a rarity rate based on the importance and class probability of each task. The appropriate rarity rate is assigned to each training sample. Thus, the samples in a mini-batch for training a deep convolutional neural network are augmented according to this rarity rate to focus on the classes with a few samples. Our heterogeneous learning approach with the rarity rate performs pedestrian attribute recognition better, especially for classes representing few training samples.

Data augmentation with rarity rate

To improve the performance for classes with few samples, we propose a method using a rarity rate assigned to each sample. The samples forming the mini-batch are chosen based on rarity rate. Note that we refer to samples from small classes as rare samples and samples from large classes as common samples. In conventional mini-batch creation, the performance for the rare classes is worse because of the random choice of training samples. The proposed method improves the performance for these classes by increasing the number of choice times that for rare samples are chosen by using the rarity rate.

This figure shows the performance of applying the proposed method when applied for VGG16. When we compare the performance of CNN, VGG16 and proposed method, VGG16 with proposed method is the best performance in an accuracy of all tasks. In the classes of low training sample, the proposed method is better accuracy than conventional VGG16 except “body orientation is back” and “face orientation is back”. In particular, “open umbrella” performance improves by more than 50% for VGG16 with the proposed method.

Bibtex