ICCV 2021 Workshop SSLAD Track 1 - 2D object detection Forum

Go back to competition Back to thread list Post in this thread

> Backbone of baseline

The backbone of baseline is ResNet50 or ResNet101?

Posted by: zrex @ Sept. 1, 2021, 11:05 a.m.

Q1:The backbone of baseline is ResNet50 or ResNet101?
R1: ResNet50 with multi-scale training, syncBN, cascaded roi head and preciseBN (when testing). Detectron2 is adopted as our codebase.

Best!

Posted by: sslad2021 @ Sept. 2, 2021, 1:32 a.m.

Do you deploy DCN or not? I cannot reimplement your baseline result without DCN base on mmdetection

Posted by: zrex @ Sept. 2, 2021, 6:11 a.m.

No, DCN is not used. As I said, we use Detectron2 as our codebase (multi-scale training (based on 1920*1080), syncBN, cascaded roi head and preciseBN (when testing)), It' maybe some difference when you use mmdetection.

Posted by: sslad2021 @ Sept. 3, 2021, 1:35 a.m.
Post in this thread