Consistent Semantic Annotation of Outdoor Datasets via 2D/3D Label Transfer

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Scene understanding models require large ground truth datasets for training and evaluation.
• Manual annotation of semantic regions in real sensor data is time-consuming and labor-intensive.

Purpose of the Study:

• To propose a framework for accelerating the semantic annotation of scenes captured by moving cameras.
• To reduce the manual effort and improve the consistency of dataset creation for machine learning.

Main Methods:

• Utilizing an existing 3D model of a scene to project segmented 3D objects into camera frames for initial 2D image annotation.
• Incorporating user-guided manual refinement of the initial annotations.
• Employing optical flow estimation to propagate refined annotations to consecutive frames.

Main Results:

• The proposed framework was evaluated during the creation of a labeled outdoor dataset.
• Annotation time was reduced by an average of up to 43% compared to traditional methods.
• Improved consistency in the semantic labeling of scene data was observed.

Conclusions:

• The developed framework significantly enhances the efficiency of semantic scene annotation for machine learning.
• It offers a practical solution for reducing the bottleneck in ground truth dataset creation for autonomous systems.
• The method improves both the speed and quality of data labeling for scene understanding applications.