Synthetic Data for 2D Road Marking Detection in Autonomous
Driving 2023-01-7046
The development of autonomous driving generally requires enormous annotated data
as training input. The availability and quality of annotated data have been
major restrictions in industry. Data synthesis techniques are then being
developed to generate annotated data. This paper proposes a 2D data synthesis
pipeline using original background images and target templates to synthesize
labeled data for model training in autonomous driving. The main steps include:
acquiring templates from template libraries or alternative approaches,
augmenting the obtained templates with diverse techniques, determining the
positioning of templates in images, fusing templates with background images to
synthesize data, and finally employing the synthetic data for subsequent
detection and segmentation tasks. Specially, this paper synthesizes traffic data
such as traffic signs, traffic lights, and ground arrow markings in 2D scenes
based on the pipeline. The effectiveness of this pipeline was verified on the
public TT100k dataset and the CeyMo dataset by image detection tasks. Template
positioning methods including random location and same position replacement were
employed for synthesis in traffic sign detection. For ground arrow marking
detection, template positioning methods encompassing inverse perspective
transformation and lane line positioning were utilized. Extensive experiments
were carried out on the TT100K dataset and the CeyMo dataset. The performance
between those open datasets and the synthetic data in this paper were then
compared. The results show that the detection model trained entirely on
synthetic data can achieve up to 86% mAP@0.5 on the TT100k dataset validation
set, and choosing 50% of the CeyMo training set for fine-tuning can achieve 77%
mAP@0.5. We have verified that data synthesis for categories with less data can
effectively mitigate the class imbalance problem in datasets. This demonstrates
that the pipeline proposed in this paper is a practical and effective approach
in the field of autonomous driving data synthesis.