Closed zzw1123 closed 5 years ago
@zzw1123 When the individual parsing results fused into the final result, the shared area of two segmentation maps will inference each other. In the case you mentioned, the shared area of the blue person and the nearby person covers the pixels around the "vertical line". For each of these pixels, the confidence on background is larger than that on foreground, leading to the phenomenon you mentioned.
@RanTaimu You mean it is because that the network first generates single person ROI whose shape is rectangle so the local parsing result is restricted in this rectangular bounding box. And then during fusing global parsing result and local result, the area outside the rectangule is considered to be "background". Am I right? So there comes another question, how to fuse them?
@zzw1123 Yes you're right. you can refer to the file 'M-CE2P-test.py' in M-CE2P for more details.
@RanTaimu Got it! Thank you so much for the help!
Thanks for your contribution! I have noticed that in the paper, some visualization results are provided in Fig.4. However, there seems like a vertical line divide the small blue person(surrounded by white bounding box) into two parts. And there is no texture difference between this two parts. I am wondering the reason.
![image](https://user-images.githubusercontent.com/16879220/50215178-2d95b900-03bd-11e9-9c6f-f4ecc0fc8fad.png)