Table 2 Frame-level performance comparisons.

From: Detection of eye contact with deep neural networks is as accurate as human experts

Metric

Mean rater

Deep model

Deep model

Deep model

Multi-task learning24

  

(smoothed)

(not smoothed)

Without transfer learning (smoothed)

With ResNet (smoothed)

F1

0.932

0.940

0.930

0.916

0.906

Precision

0.918

0.936

0.924

0.917

0.924

Recall

0.946

0.943

0.937

0.915

0.890

  1. Performance without smoothing (fourth column) is reported at the maximum F1 score along the PR curve, with associated PR. In the sixth column, we replaced AlexNet in ref. 24 with ResNet for a fair comparison.
Back to article page