search

Katie0723 / DQ-DETR

28 stars 1 forks source link

Question about the COCO experiments #3

Closed HaisongDing closed 9 hours ago

HaisongDing commented 3 weeks ago

In your paper, DQ-DETR yieldes slightly lower performance compared with DINO. My followup questions are:

  1. Is the degradation caused by the "Dynamic Query Selection (DQS)" part?
  2. Have you tried applying "Categorical Counting (CC)" and "Feature Enhancement (FE)" designs only and checked whether it will achieve better results than DINO?
hoiliu-0801 commented 1 day ago
  1. Our paper focuses on tiny object detection, with components specifically designed for this task. Achieving state-of-the-art performance across all datasets is challenging. In contrast, DQ-DETR outperforms DINO significantly in tiny object datasets.
  2. Thank you for the suggestion. We are still working on it.