The labeling service facilitates the creation of a dataset to train self-driving AI. This semi-automated labeling function uses a deep neural network algorithm to extract candidates from image data included in an input ROSBAG file. The extracted objects are labeled and added to the training dataset. These files are later manually labeled by Automan staff. However, be aware that manual labeling is not immediately available as the automated semi-labeling.
In a future patch, GUI tools will allow to modify labels in the training set to improve the accuracy of the trained model. The tool will be provide enhanced support to label 3D datasets obtained from a LiDAR sensor.
The training service enables online training of deep learning for self-driving AI based on the labeled dataset. The alpha version of this service supports the Single Shot Multibox Detector (SSD) algorithm, which is a well known deep neural network algorithm for object detection. You will obtain a Caffe compatible model, and the training parameters used once the online training finishes. These can be loaded in Autoware, a well-established open self-driving framework.
In the future we plan to support more architectures and improve the training flexibility. To further optimize performance and training time, we plan to introduce academic supercomputers and cloud GPU clusters as a backend engine of computation.
The following is an example of improvement in deep neural networks through Automan's labeling service.
Premium accounts can access to a broad range of labeled datasets provided by Tier IV.