search

openseg-group / openseg.pytorch

The official Pytorch implementation of OCNet series and SegFix.
MIT License
1.18k stars 140 forks source link

about segfix #79

Open wangdefour opened 2 years ago

wangdefour commented 2 years ago

Hi,Do you have instructions for segfix?

hsfzxjy commented 2 years ago

Please check https://github.com/openseg-group/openseg.pytorch/blob/master/MODEL_ZOO.md#segfix-1

wangdefour commented 2 years ago

@hsfzxjy Hello, I have some questions. I run segfix directly using the "offset_semantic. Zip" file. I found that my label must have a corresponding file in "offset_semantic. Zip", as shown in line 83 of segfix.py:

def get_ offset(basename): return io.loadmat(osp.join(offset_dir, basename + '.mat'))['mat'].astype(np.float32).transpose(2, 0, 1) * args.scale

If I want to generate an offset file, I must use the cityscapes format. My data is [0, 255] Does this mean that I need to retrain the offset model?

I3aer commented 2 years ago

Segfix corrects minor errors on the boundary as long as the segmentation is reliable for other pixel locations. Therefore, in my view, it is a useless network. The most important issue is how well your segmentation network is generalized. Most segmentation datasets are small and probably large networks like OCR memorize the noise in datasets instead of learning. Instead of Segfix use an active contour segmentation or CRF.

Edit: Grammar