wenhe-jia / detectron-cascade-rcnn-exp Goto Github PK

Hi LeoJWH,

From your code fast_rcnn.py in the roi_data directory, it seems you are using the same proposals from RPN in 3 stages of cascade_rcnn. But from the paper, the author said proposals of stage 2 are refined from those of stage 1 and proposals of stage 3 are from those of stage 2. Could you explain that? Thanks. @LeonJWH

请问你对于原版detectron具体在哪里进行了修改呢？

上网查了一下，说我python环境有问题，但是我在用原来的detectron子文件仍然可以训练，如何解决这个问题呢？

hi LeonWJH,
Good jobs.
I am implementing cascade rcnn recently on pytorch which follows the implementation architecture of Detectron, but I have encountered some problems. The training curves looks norma. The accuracy of the second stage is higher than that of the first stage, and the loss is lower. But when I tested, the second stage was much worse than the first stage. Do you have some ideas?