Autonomous driving
Embedded Files
PiShield can be used for real-world applications such as autonomous driving, where the models' predictions must be compliant with the background knowledge.
Task: multi-label classification, which requires training a model to predict a set of positive labels for each detected bounding box.
Possible labels: 41 labels, where 10 labels correspond to agents, 19 labels to actions and 12 labels to locations.
Training Data: video frames from the ROAD-R dataset.
Requirements: the logical constraints available in ROAD-R (i.e. 243 requirements written over 41 labels).
Example requirement: a traffic light cannot be red and amber at the same time.
Before
Before
The predictions violate the background knowledge, since the Traffic Light is assigned both Red and Amber at the same time. Such predictions can results in the self-driving vehicles making dangerous decisions.
After
After
Guaranteed satisfaction of the background knowledge.
Page updated
Google Sites
Report abuse