COCOR: Training and Assessing Rotation Invariance in Object and Human (Pose) Detection Tasks

Abstract

The performance of neural networks on human (pose) detection has significantly increased in recent years. However, detecting humans in different poses or positions, with partial occlusions, and at multiple scales remains challenging. The same conclusion arises if we consider object detection tasks. In the context of this research, we focus on the rotation sensitivity in object detection and in human (pose) detection tasks for state-of-the-art neural networks. To the best of our knowledge, there are few corpora dedicated to the rotation problem and, for people detection, to fall or fallen person detection, but none contain all rotation angles of the image that could be used to train or evaluate machine learning systems towards rotation invariance. This research proposes two variants of the COCO dataset. COCOR is a rotated version of the standard COCO dataset for object and human (pose) detections while COCOR-OBB provides oriented bounding boxes information as people annotation. The implementation details concerning the construction of COCOR and COCOR-OBB are depicted in this article. Providing baseline evaluation of SOTA systems, COCOR can be used as a benchmark dataset for rotation invariance evaluation in vision tasks, including object detection and human (pose) estimation.