Older Model (0.8.13) VOACC drastically overestimates vLead and dRel #32169
Replies: 1 comment
-
Fixed it. Didn't realize initially the comma2k19 dataset used the eon. OpenCV2 warp function from the old webcam functionality fixed it. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi all,
First of all this is a purely academic question. I've been using an old vision model from the CommaTwo_Master branch. I've been doing a lot of testing with Openpilot in CARLA. I've noticed that the lead vehicle behavior is VERY aggressive in CARLA. It tends to react to lead vehicles very late and slam on the brakes. This is due to large overestimates in dRel and vLead from the vision model. Initially, I just chalked this up to a domain shift problem. Seeing CARLA images differently than it sees real images.
However, more recently I've been playing around with the Comma2k19 dataset. When I put these images through the model, the estimates actually get even worse. A random example:
The model seems to be off by a consistent factor of around x1.5 - x2 for dRel. I'm really just curious why. Similar observation was made here for a much newer model #26799. Obviously I could adjust it by some factor, as was done in that discussion, but I'm more curious if these wide error margins are considered normal or expected, or if this indicates I've made a mistake in my implementation. In both CARLA and the 2k19 data, I've run through the calibration procedure. Lateral output from the model looks great in both cases.
Any thoughts or gotcha's would be appreciated.
Beta Was this translation helpful? Give feedback.
All reactions