Fig 1.
The multi-resolution stage modules are marked with blue color areas. The remained three stages module consists of parallel multi-resolution subnetworks with multi-resolution information interactions.
Fig 2.
The channel attention mechanism.
The top block (a) is ECA block, which consists of average pooling, a fast 1D convolution of size k and a sigmoid activation. The bottom block (b) is SE block, which consists of average pooling, two fully-connected (FC) layers and a sigmoid activation.
Fig 3.
The dySample and the Detail Enhancement Module (DEM) are applied to the HRNet [5,9], further implementation details are provided in Section 3.1.
Fig 4.
The structure of Detail Enhancement Module.
The module is designed based on SE block. Additionally, the global average pooling and dropout [25] technology are located before and after the SE block, respectively. Dashed lines denote selective identity mappings.
Table 1.
Comparisons on the COCO validation set. #Params and FLOPS are calculated only for the method.
Table 2.
Comparisons on the COCO test-dev set. The bottom-up methods use multi-scale testing.
Table 3.
Comparisons of PCK@0.5 score on the MPII valid set.
Table 4.
Ablation study of DE-HRNet’s components on the MPII valid set.
Table 5.
The ECA block and SE block components on the MPII valid set.