Skip to main content

Table 2 MPJPE performance comparison on the validation and test dataset. The table shows some parts of joints in detail, however the All includes other joints (head, neck, torso, and waist). Lower value is better in this metrics

From: 3D human pose estimation model using location-maps for distorted and disconnected images by a wearable omnidirectional camera

Model Backbone In-scale Out-scale Var. #Params GFLOPs Shoulders Elbows Wrists Hands All
Validation dataset
VNect ResNet50 96 x 192 12 x 24 1.0 14.5 M 1.70 0.019 0.053 0.069 0.079 0.044
VNect ResNet100 96 x 192 12 x 24 1.0 33.5 M 2.97 0.019 0.052 0.067 0.076 0.043
Ours HRNet-W24 96 x 192 24 x 48 1.0 16.7 M 1.70 0.015 0.044 0.060 0.069 0.037
Ours HRNet-W32 96 x 192 24 x 48 1.0 29.3 M 2.70 0.015 0.045 0.062 0.069 0.038
Test dataset
VNect ResNet50 96 x 192 12 x 24 1.0 14.5 M 1.70 0.040 0.218 0.387 0.439 0.201
VNect ResNet100 96 x 192 12 x 24 1.0 33.5 M 2.97 0.037 0.215 0.379 0.434 0.200
Ours HRNet-W24 96 x 192 24 x 48 1.0 16.7 M 1.70 0.040 0.215 0.395 0.467 0.204
Ours HRNet-W32 96 x 192 24 x 48 1.0 29.3 M 2.70 0.036 0.202 0.393 0.463 0.201
  1. The best performance estimations are italicized
\