Draw-n-Replace: A Novel Interaction Technique for Rapid Human-Correction of AI Semantic Segmentation.

A large volume of precise semantic segmentation data is required for training AI systems such as autonomous vehicles. To combine the speed of machine segmentation and the accuracy of human labeling, researchers have explored the human-AI pipeline, where the machine proposes segmentations and the human makes corrections. Unfortunately currently, machine-segmentation is too inaccurate and time-consuming for humans to correct. As a result, the majority of precise semantic segmentation is currently done manually. For the human-AI pipeline to be viable, improvements in both AI segmentation accuracy and human-correction speed are needed. We focus on the latter and present a novel interaction technique called Draw-n-Replace for faster human-correction of machine-segmentation mistakes. We show that our technique is 2.36 times faster than the conventional approach.