2020年1月5日星期日

Image Segmentation: Summary of Problems

1. Number of categories (label_nc / num_class)


Errors include:
RuntimeError: cuDNN error: CUDNN_STATUS_EXECUTION_FAILED

And other weird errors

Causes and solutions:
Imresize the label image

The default bilinear interpolation method almost inevitably generates new pixels, causing more classification problems and causing errors.

Solution: 
label = cv2.resize (label, (256,256), interpolation = cv2.INTER_NEAREST)

Use INTER_NEAREST interpolation.

Note: The method to view the number of labels: length (np.unique (label))

To be continued

Image segmentation / super pixel / parsing Github:

Face segmentation: https://github.com/zllrunning/face-parsing.PyTorch
Traditional method super pixel: https://github.com/mingyuliutw/EntropyRateSuperpixel
Generate real scene from segmentation graph: https://github.com/NVlabs/SPADE
Deeplab v3 + Pytorch version: https://github.com/jfzhang95/pytorch-deeplab-xception
CSAILVison Pytorch semantic segmentation: https://github.com/CSAILVision/semantic-segmentation-pytorch

没有评论:

发表评论