search

xiuqhou / Salience-DETR

[CVPR 2024] Official implementation of the paper "Salience DETR: Enhancing Detection Transformer with Hierarchical Salience Filtering Refinement"
https://arxiv.org/abs/2403.16131
Apache License 2.0
140 stars 8 forks source link

May I ask why didn't you compare your model with Co_DETR: DETRs with Collaborative Hybrid Assignments Training #31

Closed trinhvg closed 3 months ago

trinhvg commented 3 months ago

Question

Dear authors,

Why did you not compare your model with Co_DETR: DETRs with Collaborative Hybrid Assignments Training? I saw you mentioned it. I am curious why there are not many papers compared with Co_DETR.

Thank you,

Additional

No response

xiuqhou commented 3 months ago

Hello @trinhvg ,

Our focus of Salience-DETR is not to improve performance, but to reduce FLOPs as well. Co-DETR is a strong detector, we cannot beat it in performance, but we have faster inference speed and less FLOPs than it.

We think that only one of performance and efficiency is comparable, and it is meaningful to compare the other one. For example, we have the same performance as DINO but faster inference speed, we may say Salience-DETR suppresses DINO. But comparing Salience-DETR and Co-DETR can't get which one is better considering both performance and efficiency.

Another reason I guess why many papers don't compare their work with Co-DETR is that Co-DETR relies on faster CNN and ATSS architectures, which limits the application in many real-world tasks. The improvement of performance may come from the ensemble of several detectors, rather than a better design for DETR. I don't think this is a good and general choice to improve DETR.

trinhvg commented 3 months ago

Thank you for sharing your insight about this line of DETR, It helps me a lot.