Automated labelling of objects as the basis of Big Data

In the automotive industry, the demands on Big Data are increasing due to the introduction of semi-autonomous driver assistance systems. As very large amounts of reference data are required for the next generation of autonomous driving functions, automated algorithms that allow to minimize the time required and to keep the costs within a real framework have to be sought.

After the manual labeling of each individual object, an automatic object annotation is performed. This means that with the help of the software, objects known to the algorithm do not have to be manually labelled repeatedly. The software is operated in the cloud in order to efficiently access the large amounts of data. In addition, the easy accessibility in the cloud facilitates project coordination and the processing of large data sets.

The customer tried to overcome this challenge by splitting the large amount of data and manual labeling via different providers abroad. 

• Different driving scenarios with different degrees of complexity (motorway vs. urban)
• Many different object classes to be detected, which hardly differ from each other
• Data management and permanent storage of huge amounts of data

• temporal flexibility
• reasonable hourly rates
• constant high quality

• A scalable deployment concept allows the tool to be easily rolled out and deployed.
• By managing data in the cloud, the data can be efficiently managed centrally and viewed from anywhere after uploading.
• By training the neural networks specifically for the customer, added value can be created in customer-specific use cases.
• Automatic recognition of customer-specific object classes.
• Time and cost savings through efficient and automatic labeling.

• The customer presents the scenario and the data to be annotated and explains which objects are to be recognized in which form. This is best done with a sample of data or exemplary data.
• The customer then receives an assessment of which automations are possible. An estimation of the amount of data required provides a reference as to how much he has to label manually before training his deep learning network is possible.
• The customer receives an instance of the tool in a cloud environment.
• Quality checks by a workforce confirm the correctness of the annotations before delivery.

• Fundamental IT know-how for ordering and setup
• Short introduction for the use of the tool
• Training experience for neural networks is advantageous - otherwise this is provided as a service by the expert

• After determining the scenario to be annotated, an instance in the cloud and the specific application was set up for the customer.
• The data was loaded into the cloud. The customer manually annotated a portion of the data.
• The network is trained and provided.
• Pre-annotation of the data.
• A quality check of the annotations is carried out, if necessary, further data is labelled and trained. This step is repeated until a satisfactory quality in object recognition is achieved.
• Delivery/use of the annotations.

Technical Facts

• Pilot phase
• Cloudbased application with docker-based deployment
• Deep Neural Networks for object detection
• Computer vision algorithms for object tracking
• Application is accessible to the user via common browsers (Google Chrome, Firefox)
• Can also be deployed in the customer's own cloud

Integration/APIs

Connection of Azure BLOB Storage Containers to the application. 
Connection of other cloud storage possible. 

Where is the data stored?

The data is stored in the cloud storage. The cloud storage is currently located in the dedicated Bertrandt Azure Cloud. However, it is also possible to connect the customer's own cloud storage.