samson braun

Iterative Labeling Process

Zhiyong Zhang, Samson Braun, Pushyami Kaveti

In this paper, we introduce a robust and cheap way to make training data set for object detection, especially for specialized fields that lack a large data set. The main idea of the Iterative Labeling Process is to train on predictions iteratively. Amazon MTurk is used to correct predictions. Auto-approval is applied to filter the MTurk results, which make the process fully automated. The process can save three times the common labeling cost. Furthermore, it can also complement missing objects and add ”background” labels in any existing data set. Train background labels can effectively reduce false positives.