Skip to content

Decouple Post-processing from PointPillars#548

Draft
kamalrajkannan78 wants to merge 1 commit intomainfrom
kkannan/mar27_decouple_pp_from_pointpillars
Draft

Decouple Post-processing from PointPillars#548
kamalrajkannan78 wants to merge 1 commit intomainfrom
kkannan/mar27_decouple_pp_from_pointpillars

Conversation

@kamalrajkannan78
Copy link
Copy Markdown
Contributor

@kamalrajkannan78 kamalrajkannan78 commented Mar 27, 2026

Ticket

Problem description

  • Pointpillars model fails with PCC=0.7147460193298709
  • Pcc drops below 0.99 after sigmoid operation but sigmoid unit test passed with PCC=1
  • This suggests the sigmoid op itself is not broken, but is amplifying small numerical differences introduced earlier in the pipeline.
  • Since the drop is observed in the post-processing path, that stage can be decoupled from the main model forward.
  • For full investigation details, see #3986

What's changed

  • Decoupled the PointPillars post-processing path from the main model forward.
  • Moved predicted bbox decoding to loader-side post-processing.

Checklist

  • verified the changes through local testing

Logs

Before fix

After fix

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.

PCC drop in pointpillars, PCC=0.7147460193298709

1 participant