Unsupervised domain adaptation in human detection for various human pose
In human detection algorithms, some recurrent issues are still challenging. Indeed, in a real application we have a lack of various human poses in our data. For example, in video surveillance it could be a significant issue to forget detecting a human in uncommon pose. We propose a compensation to these less frequent cases. We introduce a detection in rotated view to compensate limited various poses in our dataset. Instead of introducing a generative enhancement for every human body pose, which could be exhausting. Indeed, data annotation has become a difficulty for many applications. Moreover, a simple data augmentation is not efficient to generalize our training. We need a strategy to learn this new poses. So, we will introduce a process to compute pseudo labels in a sequence of rotated views. It avoids to use a clustering algorithm like dbscan[3] or K-means to compute a pseudo label. A summary of theoretical results is described in a first section to illustrate our baseline. First of all, we will illustrate how we process pseudo label on various views by two mechanisms which have been called relaxation and scanning. We introduce a sequence of rotations of original view, to process a detection in a compensated view. In second step, we will learn this distortion created by the various views and construct an approach to distill original knowledge to a more generalized training. The main improvement of this approach is to propose an alternative solution at dbscan[3] approach which is popu- lar, but which is difficult to use in real applications. We ob- tain a significant result by improving accuracy in especially various pose datasets.