A Semi-Automatic Annotation Technology for Traffic Scene Image Labeling Based on Deep Learning Preprocessing

摘要

Massive traffic scene data for algorithm research and model training is the fundamental for self-driving car technology development. In the procedure of scene image labeling, the most accurate method is manual annotation, but with the increasing of the amount of image data, artificial annotation method becomes infeasible due to its disadvantages of vast cost, inefficiency and subjective deviation. In order to reduce the time and cost of annotation processing while ensuring the accuracy, this paper proposed a semi-automatic annotation procedure. The core idea is using an automatic preprocessing method developed with convolution neural network to roughly annotate the image, just before the human review and revision. The innovation of the proposed automatic annotation method includes: 1.) the results of CNN processing are improved by combining with the object detection outcome; 2.) a variable parameter outlier-merging algorithm based on the sliding window is provided to deal with the large number of outliers. It shows about 5 percentage improvement in class average accuracy and 4/5 decrease of time to process one picture by using our processing method.