End-to-End Learning for Monocular 3D Human Pose Estimation

3D human pose estimation (3DHPE) based on a single RGB image is a fundamental problem in the field of computer vision. The estimation accuracy and speed of 3DHPE directly affect the practical applicability. However, existing methods often improve accuracy by using complex network architectures or multi-stage processing. These approaches result in more computational overhead and slower estimation speed. To balance estimation accuracy and speed, we need a more efficient approach. Motivated by the need for both estimation accuracy and speed, we propose the I-KDnet, an end-toend model, that achieves precise 3D pose estimation at a high speed. Specifically, we design an Idealized Knowledge Distillation (I-KD) training approach, an idealized variant of online knowledge distillation. During training, I-KD enhances the training process, similar to online knowledge distillation. During inference, it does not introduce any additional computational overhead. Additionally, compared to online knowledge distillation, the I-KD approach is easier to implement and more effective. Based on this approach, I-KDNet achieves the best accuracy on the Human3.6M benchmark with high estimation speed.