fcis/demo.py classifies the entire image as 50+ examples of fire hydrant

msracver / FCIS

Fully Convolutional Instance-aware Semantic Segmentation

MIT License

1.57k stars 414 forks source link

fcis/demo.py classifies the entire image as 50+ examples of fire hydrant #76

Open cbra12 opened 7 years ago

cbra12 commented 7 years ago

Hi, I am attempting to get the fcis/demo.py code up and running. I have followed the instructions for installation and when I run the demo.py script the test images are oversegmented into 50+ classes of fire hydrant covering the entire image. I can't seem to find anyone else having this issue, does anyone have any suggestions? Below are two of the outputs for the demo. Any help would be greatly appreciated! demo1 demo2

chrieke commented 7 years ago

The different colors don't indicate different classes, that is just so that you can differentiate them. They are all proposals of the fire hydrant class. Usually it would display all proposals with a detection score treshhold > x, e.g. 0.7 etc (which would usually filter out pretty much all proposals that don't make any sense). But all proposals here seem to have 1.00, which is pretty weird.

No idea what went wrong. Have you doublechecked and installed everything from scratch as in the readme?

liyi14 commented 7 years ago

I met this situation before, but only remember it is due to my careless skipping of some steps during installing. Maybe you can read the README again and see which steps if skipped will leads to this problem.

mldm4 commented 6 years ago

I was in the same situation and I solved it by updating to CuDNN v5 (I was using v4).