RevPose: A Multi-Stage Reversible Networks for Human Pose Estimation

We propose a novel multi-stage reversible network for human pose estimation, which is called RevPose. The proposed RevPose consists of network modules with a variable number of columns that can be continuously increased. Multi-column reversible connection approach is employed in each sub-network. This approach is distinct from existing traditional networks, as it gradually unfolds the learned information during the forward propagation process across multiple stages while avoiding operations that can lead to loss of feature, such as fusion and downsampling. Our experiments show that the RevPose achieves fascinating performance. For example, with no pre-train, RevPose get 77.3% AP on the human pose estimation on the COCO dataset. It prove the RevPose, as a new try of reversible network, exhibits superior replaceable properties and can be used as a new direction for future research.