What we do
To make autonomous driving as safe as possible we evaluate data for neural networks.
Testing data can be images or LiDAR point cloud maps. Our aim is to make sure that the neural network is trained with almost 100% flawless data. To reach this goal we write tools that help us to evaluate more precisely.
2D
Semantic Segmentation
Aims to make the class prediction for an entire input image. To reach this goal each pixel is labeled with the specified color of its object.
Full Image Tagging
Full image tags provide general information about an input image such as environment and weather conditions.
Bounding Boxes
Rectangles to enclose a specific object or region. Provides tracking of different classes like cars, pedestrians or signs.
Object Tagging
Object tags provide specific information about an input image. Every object has its own requirements that need to be considered.
Skeleton Annotations
Visualizes human movements for a static input image by annotating different body parts. Is used for predicting motion sequence of a pedestrian.
Road Model Segmentation And Tagging
Aims to make the lane prediction for an entire input image. To reach this goal each lane is labeled with the specified color.
3D
Semantic Segmentation
Aims to make the class prediction in 3D LiDAR space. To reach this goal each point is labeled with the specified color of its object.
Bounding Boxes And Object Tagging
Cuboids to enclose a specific object or region in 3D LiDAR space. Provides tracking of different classes like cars, pedestrians or signs. Object tags provide specific information about a LiDAR cloud. Every object has its own requirements that need to be considered.
A few things we’re great at
There are a few things that makes us special. At the very beginning we made sure that we can work with different companies in the future.
FLEXIBLE TOOL
Due to our good structured code of our evaluation tools we are prepared for changing requirements. Each company has its own practices we can handle.
CONCENTRATET EVALUATORS
We know how important it is to feed neural networks with correct data. A neural network can only be as good as the training data is. Autonomous driving is a very sensitive topic, we definitely want to make sure it is fed with verified correct data.
POWERFUL TESTING AREA
With just a few clicks we can detect incorrect data without even looking at them. Nevertheless we verify the data to enhance the correctness.
YEARS OF EXPERIENCE
We have experience in evaluating data since 2015. In these years we improved the methods to verify data more effectively.
Evaluation
2D Bounding Box
Object Tagging
Semantic Segmentation
3D LiDAR Bounding Box
Automatic Tests
Pre Evaluation
| Testname: | Relevance: | Message: | Details: |
|---|---|---|---|
| 2DBBoxSpecification | ERR | BBox missing | object is missing a bounding Box |
| 2DBBoxSpecification | ERR | BBox missing | object is missing a bounding Box |
| 2DBBoxSpecification | ERR | BBox missing | object is missing a bounding Box |
| 2DBBoxSpecification | ERR | BBox missing | object is missing a bounding Box |
| 2DBBoxSpecification | ERR | BBox missing | object is missing a bounding Box |
| 2DBBoxSpecification | ERR | BBox missing | object is missing a bounding Box |
| 2DBBoxSpecification | INFO | 2DBBoxSpecificationTest failed | Errors occured: 6 Tested: 3 files from: 3 |
Explanation: This is an automatic test that can be executed right before the evaluation. With intelligent methods of this tool you can find out the most trivial mistakes like missing bounding boxes, as you can see in the table above.
Runtime
| Testname: | Relevance: | Message: | Details: |
|---|---|---|---|
| 3DBBox | WARNING | ObjectTag mismatch | 13 not contained in the 3D space |
| 3DBBox | WARNING | ObjectTag mismatch | 14 not contained in the 3D space |
| 3DBBox | WARNING | ObjectTag mismatch | 12 not contained in the 3D space |
| 3DBBox | WARNING | ObjectTag mismatch | 11 not contained in the 3D space |
| 3DBBox | WARNING | ObjectTag mismatch | 9 not contained in the 3D space |
| 3DBBox | WARNING | ObjectTag mismatch | 8 not contained in the 3D space |
| 3DBBox | WARNING | ObjectTag mismatch | 16 not contained in the 3D space |
| 3DBBox | INFO | 3DBBox succeeded | Tested: 3 files from: 3 |
Explanation: This is an automatic test that is executed during the runtime. With intelligent methods of this tool you can find out the most trivial mistakes like non-matching object tags, as you can see in the table above.
