DJW cdece0b32a 第一次提交 | 10 月之前 | |
---|---|---|
.. | ||
coco | 10 月之前 | |
README.md | 10 月之前 |
Top-down methods divide the task into two stages: object detection, followed by single-object pose estimation given object bounding boxes. At the 2nd stage, integral regression based methods use a simple integral operation relates and unifies the heatmap and joint regression differentiably, thus obtain the keypoint coordinates given the features extracted from the bounding box area, following the paradigm introduced in Integral Human Pose Regression.
Results on COCO val2017 with detector having human AP of 56.4 on COCO val2017 dataset
Model | Input Size | AP | AR | Details and Download |
---|---|---|---|---|
ResNet-50+Debias-IPR | 256x256 | 0.675 | 0.765 | resnet_debias_coco.md |
ResNet-50+DSNT | 256x256 | 0.674 | 0.764 | resnet_dsnt_coco.md |
ResNet-50+IPR | 256x256 | 0.633 | 0.730 | resnet_ipr_coco.md |