Update README with LTDETRv2 m/l benchmark results#894
Conversation
s-coco was the only LTDETRv2 size with published COCO numbers. Add m and l (53.1 / 56.0 mAP) with T4 latency from the LT-DETRv2 latency benchmark; x is still training so its mAP is marked TBD. Checkpoint uploads for m/l/x are tracked separately.
gabrielfruet
changed the title
|
/codex review |
|
/review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 92bcdae767
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
yutong-xiang-97
left a comment
yutong-xiang-97
left a comment
There was a problem hiding this comment.
We can add a plot here and compare with the competitor models!
| @@ -95,6 +93,8 @@ Train LTDETR detection models with DINOv2, DINOv3, or EdgeCrafter ECViT backbone | |||
| | picodet-s-coco | 26.7\* | 2.2\* | 1.17 | 416×416 | | |||
There was a problem hiding this comment.
I would actually want to drop picodet numbers or move it to the bottom but I will leave it to you to decide.
There was a problem hiding this comment.
Or alternatively, group the results with model backbones. I think up to this point the table is a bit too long to read.
There was a problem hiding this comment.
I think it makes more sense to drop picodet
2 participants
Depends on: #887 (must merge first or together — the m/l
-cocoaliases don't exist until then)What has changed and why?
ltdetrv2-s-cocowas the only LTDETRv2 size with published COCO numbers inthe README. Adds
ltdetrv2-m-coco(53.1 mAP) andltdetrv2-l-coco(56.0 mAP),plus T4 latency for all three sizes from internal benchmarking.
Checkpoint uploads for m/l (wiring the
-cocoaliases to downloadablecheckpoints) are tracked in a separate PR — this one only updates the
benchmark table.
How has it been tested?
README-only change; visually checked the rendered table for alignment.
Did you update CHANGELOG.md?
Did you update the documentation?