Po-Chih Huang

Generally, when it comes to IR ranking with deep learning model, we use the outputs of the second last layer as input image's feature. Because the outputs of the last...

You can create your own dataset from web.

You have to make sure the pixel-wise label exists for you picture, so it can be trained.

 **701_StillsRaw_full** is a directory of original images. **LabeledApproved_full** is a directory of labeled images. **label_colors**.txt is a file which told the label of a color.

There are lots of reason to cause this. 1. Deep learning technique is not stable. 2. The implemented method in this repo is old, so the miou isn't competitive. 3....

You mean the results? You can save the output of pre-trained model.

Can you report the dataset you used? Thanks!

Which dataset did you run

IoU is about 0.65~0.70

The output and the target are both construct with 0 and 1 value. So it's trainable with BCELoss.