Local and Global Orientation Correction for Oriented Human (Pose) Detection
Résumé
Detecting people and/or their pose in real life condition is of interest for several tasks notably for home care for elderly or frail people for instance. In such contexts, the perception system must be able to detect usual but also unusual poses of people in various adversarial conditions. Even if the performance of neural networks on human (pose) detection has significantly increased recently, the human detection in different poses or positions, with partial occlusions, and at multiple scales remains a challenge. In this research, a step towards rotation invariance of human detection is proposed, i.e. to identify a person in a robust way on images containing rotated or oriented human poses. After confirming that data augmentation could not solve
this problem, this research explores three ways to address the rotation problem in human pose detection: steerable networks, global rotation correction, and local person orientation approaches. From reported experiments on rotated generated corpora from COCO dataset, there is no significant improvement from integrating steerable approaches into existing architectures. While the state-of-the-art approaches lose up to 67.2 mAP
on rotated images, the global rotation correction keeps almost intact the performance for all angles but does not solve the problem of images crowded with several people at various orientations. The local orientation approach permits an average of 4.3 mAP[.5:.95] and 9.5 mAP0.5 gain, and even more when combined with the global approach. The reported experiments indicate that it is possible to achieve rotation invariance. The paper ends by discussing possible improvements to strengthen the rotation invariance in perceiving humans.