search

SHI-Labs / Decoupled-Classification-Refinement

Revisiting RCNN: On Awakening the Classification Power of Faster RCNN (ECCV 2018)
https://arxiv.org/abs/1803.06799
MIT License
167 stars 24 forks source link

Question about Boxes used in RCNN #13

Closed JunweiLiang closed 5 years ago

JunweiLiang commented 5 years ago

Hi, do you use boxes from RPN (so usually is (1000, 4)), or boxes after faster-RCNN box regression (also (1000, 4)), or the final boxes after NMS?

bowenc0221 commented 5 years ago

I use boxes after faster-RCNN box regression before NMS.

JunweiLiang commented 5 years ago

Minor question: so in Fig.4 of your paper, there shouldn't be an arrow from RPN proposals pointing to RCNN training? What does that arrow mean?

bowenc0221 commented 5 years ago

The main reason is to make the figure looks clean without overlapping arrows @weiyc It should come from the Faster R-CNN outputs as stated in Section 4.2.

JunweiLiang commented 5 years ago

Thanks. Have you tried using RPN box instead? If you weren't using class-agnostic regressor head then the regression output depends on the Faster RCNN classification output, which I think may negatively affect the Decoupled RCNN performance.

bowenc0221 commented 5 years ago

The bounding box regression in Faster R-CNN is class agnostic. That is, all classes shared the same box for each proposal.