SJTU-Thinklab-Det / r3det-on-mmdetection

R3Det: Refined Single-Stage Detector with Feature Refinement for Rotating Object
Apache License 2.0
177 stars 42 forks source link

Feature refinement module (FRM) points #11

Open Hussni-V opened 4 years ago

Hussni-V commented 4 years ago

According to the paper, the FRM samples five vectors per feature point during feature refinement, However, after a closer look in the code I found out that it is only sampling the center point vector and not the other 4 corners. This setting cannot be changed through the config and is not obvious without looking deep into the code.

I just want to ask why the default setting is as such and not as said in the paper?

DayBreak-u commented 3 years ago

I see the code with fr: scalar_t roi_y = bbox_offset[0] spatial_scale; scalar_t roi_x = bbox_offset[1] spatial_scale;

scalar_t px[5] = { roi_x, 0, 0, 0, 0 }; scalar_t py[5] = { roi_y, 0, 0, 0, 0 };

why roi_y is bbox_offset[0] not bbox_offset[1]?