DetectoRS with mmdetection2

joe-siyuan-qiao / DetectoRS

DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution

Apache License 2.0

1.14k stars 176 forks source link

DetectoRS with mmdetection2 #75

Open CuongNN218 opened 4 years ago

CuongNN218 commented 4 years ago

Hi, Thanks for your great work but I really confuse when looking at your config files in DetectorS htc in v2 of mmdetection. In mmv2, you inherited htc_r50_fpn and it have FusedSemanticHead with 183 classes. Why it 183 classes? if i want to use you cfg file with one class i can change it to 1 ?

joe-siyuan-qiao commented 4 years ago

HTC has an additinoal semantic segmentation branch, hence 183 classes. You can set base to htc_without_semantic_r50_fpn_1x_coco.py which doesn't require those 183 classes.

CuongNN218 commented 4 years ago

@joe-siyuan-qiao hi, i use mmdetection2 with detectors resnext 101 and train from scratch (no pretrained) but I got nothing, loss value is quite good ( i had set grad clip with ) but testing result is always nothing. I have try cascade fastercnn x101 and the result is quite good. Most of objects in my dataset account for a small area in picture. How can i deal with it?

CuongNN218 commented 4 years ago

My bactchsize is 4