search

NVIDIA / pix2pixHD

Synthesizing and manipulating 2048x1024 images with conditional GANs
https://tcwang0509.github.io/pix2pixHD/
Other
6.63k stars 1.39k forks source link

How to generate label maps for training our own dataset? #166

Open MuhammadAsadJaved opened 4 years ago

MuhammadAsadJaved commented 4 years ago

Hello everyone @all

I want to generate infrared images from visual images.

I have visual and corresponding infrared images and labels in xml files. 1- which method is better to train using these visual and infrared images to generate new infrared images? 2- how to generate label maps from existing visual or infrared images to train using label maps?

3- any other suggestions for this task?

Note: I have tried pix2pix using pair images before but the results are not very good. The objects are not clear.

kail85 commented 4 years ago

label maps are simply segmentation images. If you don't have them you may manually annotate them or using some algorithm to do the segmentation.

MuhammadAsadJaved commented 4 years ago

@kail85 I have tried keras image segmentation but failed to get segmented images. any other project suggestion?

Thank you

kail85 commented 4 years ago

Have you tried Otsu's method in OpenCV?

MuhammadAsadJaved commented 4 years ago

@kail85 Not yet. I will give it a try. I will be thankful if you share any link. It will save my time.

anant-patankar commented 3 years ago

@kail85 Not yet. I will give it a try. I will be thankful if you share any link. It will save my time.

You can use UNet for semantic segmentation. First annotate some of images with LabelMe or CVAT annotation tool manually and then train UNet with it and Make predictions on other un-annotated images

SherryWu178 commented 2 years ago

I wonder does label maps produce better results compared to RGB images?