Use Case and High-Level Description

This is a multi-person 2D pose estimation network based on the EfficientHRNet approach (that follows the Associative Embedding framework). For every person in an image, the network detects a human pose: a body skeleton consisting of keypoints and connections between them. The pose may contain up to 17 keypoints: ears, eyes, nose, shoulders, elbows, wrists, hips, knees, and ankles.



Metric Value
Average Precision (AP) 54.3%
GFlops 14.3253
MParams 8.1506
Source framework PyTorch*

Average Precision metric described in COCO Keypoint Evaluation site.


Image, name: input, shape: 1, 3, 448, 448 in the B, C, H, W format, where:

  • B - batch size
  • C - number of channels
  • H - image height
  • W- image width

Expected color order is BGR.


The net outputs are two blobs:

  1. heatmaps of shape 1, 17, 224, 224 containing location heatmaps for keypoints of all types. Locations that are filtered out by non-maximum suppression algorithm have negated values assigned to them.
  2. embeddings of shape 1, 17, 224, 224, 1 containing associative embedding values, which are used for grouping individual keypoints into poses.

Legal Information

[*] Other names and brands may be claimed as the property of others.