Table 5 Mean OKS score for selected best-performing models

From: Neonatal pose estimation in the unaltered clinical environment with fusion of RGB, depth and IR images

Backbone

Model

Uncovered

Half-Covered

3/4 Covered

Fully Covered

HRNet-W32-256

Depth

0.925

0.898

0.880

0.753

HRNet-W32-256

IIF-2

0.933

0.915

0.892

0.772

HRNet-W32-384

LIF-3

0.938

0.908

0.882

0.805

HRNet-W48-384

RGB

0.923

0.919

0.877

0.695

HRNet-W48-384

LIF-3

0.941

0.916

0.882

0.794

HRFormer-S

IIF-2

0.934

0.910

0.892

0.752

HRFormer-B

IIF-2

0.933

0.918

0.891

0.780

Backbone

Model

Prone

Supine

Side

Intervention

HRNet-W32-256

Depth

0.935

0.934

0.823

0.837

HRNet-W32-256

IIF-2

0.952

0.929

0.836

0.861

HRNet-W32-384

LIF-3

0.932

0.946

0.850

0.877

HRNet-W32-384

RGB

0.933

0.919

0.805

0.846

HRNet-W48-384

LIF-3

0.940

0.942

0.843

0.889

HRFormer-S

IIF-2

0.938

0.930

0.824

0.866

HRFormer-B

IIF-2

0.944

0.940

0.835

0.866

  1. Each row shows the score for each level of covering or position. For every model, the score decreases as the covering increases. There is no general trend between prone and supine, but every model scores lower during interventions than prone/supine and lower still when in the side position. Note that mean OKS scores are typically higher than the AP value.
Back to article page