Simulated Intersection Scenarios for Video Generation Evaluation
We evaluate our approach using simulated intersection scenarios on NuScene data, where the ego car (green) remains stationary. In the bird's eye view map, we manually place new agents (red) crossing the intersection from left to right or right to left. These agents' bounding boxes are then projected onto the ego car's perspective. The resulting bounding-box frames are used as input for the Box2Video network, which generates realistic video sequences.
Our model has not been trained on car crash examples. To evaluate its generative capabilities on out-of-distribution scenarios, we tested it using manually crafted car crash prompts. In this scenario, the ego car (represented as a green car in the bird's-eye view map) remains stationary. Two new agents (red) were positioned to cross the intersection from opposite directions, leading to collision between them. These agents' bounding boxes are then projected onto the ego car's perspective. The resulting bounding-box frames are used as input for the Box2Video network, which generates realistic video sequences. Upon collision, the two cars merge into one in the output video, illustrating the model's response to an out-of-distribution input.
