Skip to content

Add instance segmentation benchmarking (RF-DETR-Seg, D-FINE-Seg, EC-Seg) - #5

Open
EHxuban11 wants to merge 1 commit into
mainfrom
add-segmentation-task
Open

EHxuban11 wants to merge 1 commit into
mainfrom
add-segmentation-task

Conversation

@EHxuban11

Copy link
Copy Markdown
Contributor
  • ModelSpec.task field; 13 seg registry keys with per-variant native sizes
  • Collect per-instance RLE masks (segm-only prediction list so pycocotools keeps mask areas for AP_small/medium/large)
  • Dual COCO eval (bbox + segm); mask mAP is the headline accuracy for segment-task submissions, box mAP kept under bbox_-prefixed keys
  • model.task recorded in submissions (additive, va.submission.v1 compatible)
  • Fail loud if a segment-task model returns no masks
  • Verified end to end: rfdetr-seg-n pytorch/cuda on val2017 subset

Claude-Session: https://claude.ai/code/session_01L6cgJ7KpTTCq3RzD6rhFJi

- ModelSpec.task field; 13 seg registry keys with per-variant native sizes
- Collect per-instance RLE masks (segm-only prediction list so pycocotools
  keeps mask areas for AP_small/medium/large)
- Dual COCO eval (bbox + segm); mask mAP is the headline accuracy for
  segment-task submissions, box mAP kept under bbox_-prefixed keys
- model.task recorded in submissions (additive, va.submission.v1 compatible)
- Fail loud if a segment-task model returns no masks
- Verified end to end: rfdetr-seg-n pytorch/cuda on val2017 subset

Claude-Session: https://claude.ai/code/session_01L6cgJ7KpTTCq3RzD6rhFJi
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