Role of object recognition considering the existence of object segmentation

[ahmedius2]

Please help me figure out a scenario that I can simulate using AWSIM where the latency and accuracy of object recognition (utilizing lidar DNN) would play an important role in autonomous driving performance. The performance could be measured by the number of collisions or time to reach a destination safely. If there are any other performance metrics, please let me know.

What confuses me is the existence of object segmentation alongside object recognition, which already provides a lot of the information needed to avoid clashes. So we need to think of cases where segmentation would not be enough. One idea I have is the scenario of turning like in the image below where object recognition with prediction information would be needed. But I see that the tracking and prediction can also be made with segmentation results.

Thanks you for your time and help!