zhengli-hub
commented
3 weeks ago Hi, @rezama , @google-admin , @alexzzhu ,and @patzm,
As a follow up on this issue, I've included a couple of exmaples to illustrate the problem.
In uncompressed_tf_example_training/training_tfexample.tfrecord-00008-of-01000, specifically the 4th scenario, I visualized the trajectory and noticed this pattern: the vehicle enters the intersection while the traffic lights are still in red. The traffic light then turns green when the vehicle has already been in the intersection. The logical sequence should be that the traffic light turns green first, and then the vehicle starts to move. This is not the only place where the misalignment occurred. Based on the results, there are a handful of the similar issues across the dataset. Let me know if you need more examples to investigate.
Thanks!
t=0 Traffic lights are in red and vehicle stopped
t=1 Vehicle starts to move while the lights are still in red
t=2 Vehicle moves while the lights turn green.
The time is relative for demonstration purpose.
I have a question about the tfexample data format in the Motion dataset V1.2.1. Are the timestamps for traffic signals synchronized with those for vehicle positions and speeds? I'm wondering if the vehicle timestamps might precede the traffic signal timestamps, which seems problematic. I've noticed some strange phenomena: For instance, when a traffic light at an intersection changes from red to green (stop state to go state), vehicles should wait until the green light before entering the intersection. However, I've observed numerous trajectories where vehicles start moving and enter the intersection before the light changes from red to green. I'm considering whether the timestamps between traffic lights and vehicles are properly aligned.