ucas-vg / PointTinyBenchmark

Point based and tiny object detection and localization code set of UCAS-VG
http://vision.ucas.ac.cn/sources
MIT License
654 stars 80 forks source link

How did you change the SSD-Detector for Deformable DETR? #87

Open ilikeokoge opened 1 year ago

ilikeokoge commented 1 year ago

Hi, thank you for the great repojitory.

I have read your paper on SSD-Det. I believe that SSD-Det is a framework based on Faster-RCNN. That is, it detects proposals and then classifies the classes. On the other hand, you also show experimental results where you changed Detector to Deformable DETR and trained it.

My simple question is, how is a Detector that does not require a proposal, such as DETR, trained? Are there any changes to the framework? Or is it just DETR between backbone-Detector and the rest is identical?

Are example implementations available in this repository?