search

jshilong / DDQ

Dense Distinct Query for End-to-End Object Detection (CVPR2023)
Apache License 2.0
244 stars 6 forks source link

Improvement over using TOOD loss for auxiliary. #16

Open jihwanp opened 1 year ago

jihwanp commented 1 year ago

Hi, thanks for providing such a wonderful work. There seems no ablation when using a better loss design for dense queries in the paper. Compared to using focal loss, how much does it improve the performance? And different from DDQ-FCN, why did you use TOOD loss only for auxiliary in DDQ-DETR?

jshilong commented 1 month ago

Sorry for the late reply, It has been a long time since I last time checked the issue of this repo because I changed the research topic to generative models last year.

  1. "using a better loss design for dense queries in the paper." is actually what they do in their original paper. We add this loss mainly to show one-to-one assignment can also achieve SOTA 1X performance on COCO.
  2. Auxiliary queries in DETR should play the same role as dense queries in one-stage detectors.