TUBS Labeling Tool

Our labeling tool is released on GitHub under the non-commercial CC BY-NC-SA 3.0 license. If you are aiming for a commercial application, please contact us. If you are using our labeling tool in your research, please cite our paper.

 

Laser scan labeling

  • LiDAR-based labeling in bird's eye view
  • Prediction of objects into subsequent samples
  • Interacting Multiple Model Kalman filter to estimate dynamics
  • Optimization algorithms (e.g. ICP) to minimize manual editing effort
  • Labeling of approx. 70 consecutive frames / hour
  • Classes configurable freely

 

Image labeling

  • Visual support using up to four cameras, e.g. by projecting rays of sight
  • Image labeling based on projection into the image planes
  • 3D bounding box projection
  • Coarse segmentation with polygons
  • Object prediction using the LiDAR-based tracking algorithm

License plate and face detection

In order to obscure personal data within our images, we have developed a Deep Learning based license plate and face detection system. Therefore, we created our own training dataset and applied the RetinaNet (an effective neural network for object detection) to our task.

Our system achieves excellent results regarding the license plate detection and an acceptable face detection performance. The system greatly reduces manual editing effort when removing personal data from images. If you are using our system in your research, please cite our paper.

Requirements

  • Python 3.6
  • TensorFlow 1.12

Download

TUBS Dataset I/O Functions

We provide a core set of functions used by our labeling tool to read and write the data format. Furthermore, we provide an example script that reads and displays samples within the TUBS Road User Dataset.

Requirements

  • MATLAB R2015b and higher

Download