Skip to content

modeld: use correct intrinsics for extra transform when wide camera disabled#38239

Open
Quantizr wants to merge 1 commit into
commaai:masterfrom
Quantizr:fix-disable-wide-road
Open

modeld: use correct intrinsics for extra transform when wide camera disabled#38239
Quantizr wants to merge 1 commit into
commaai:masterfrom
Quantizr:fix-disable-wide-road

Conversation

@Quantizr

Copy link
Copy Markdown
Contributor

Description

DISABLE_WIDE_ROAD flag currently leads to using the ecam intrinsics on the narrow cam for the big_img model input. This PR uses the proper intrinsics if the wide camera is not available.

vlcsnap-2026-06-24-17h36m39s456

@Quantizr

Copy link
Copy Markdown
Contributor Author

Ran some tests to see how close the model path is to original given the weirdly warped big_img, this change leads to a consistent noticeable improvement. The old version still drives well enough that I've been using it for almost a year as the current models barely look at the big_img (in fact, if you just put a static gray for big_img the model drives alright). On the other hand, only using the wide cam and cropping for the narrow cam with DISABLE_ROAD makes the driving almost unusable.

Time-weighted lateral path deviation in meters vs typical two camera setup across 38 segments:

# scenario mean dev
1 DISABLE_WIDE_ROAD 0.049
2 WIDE = NARROW 0.050
3 GRAY128_WIDE 0.060
4 FIXED DISABLE_WIDE_ROAD 0.037
5 DISABLE_ROAD 0.295

@github-actions

Copy link
Copy Markdown
Contributor

Process replay diff report

Replays driving segments through this PR and compares the behavior to master.
Please review any changes carefully to ensure they are expected.

✅ 0 changed, 66 passed, 0 errors

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant