How to train with a dataset similar to Cityscapes

[chandlerbing65nm]

Is there a way to train self-supervised models here with a dataset similar to Cityscapes?

Below is an example of Cityscapes dataset. The right image is train image and the left is the annotation.

I have RGD remote sensing dataset formatted similar to this where there are training and annotation folder, usually used for semantic segmentation. However, I wanted to train it first in self-supervised method before doing the actual segmentation training in mmsegmentation.

[fangyixiao18]

You can try MAE with your datasets, which is a baseline method recently on ImageNet, but not sure about the performance on your own remote sensing dataset.

[chandlerbing65nm]

You can try MAE with your datasets, which is a baseline method recently on ImageNet, but not sure about the performance on your own remote sensing dataset.

You mean this? Masked Autoencoders Are Scalable Vision Learners

[fangyixiao18]

yes

[chandlerbing65nm]

It also says here that I can use .tif image format as input:

mmselfsup/mmselfsup/datasets/data_sources/imagenet.py

Line 71 in 684875d

IMG_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif')

mmselfsup/mmselfsup/datasets/data_sources/imagenet_21k.py

Line 36 in 684875d

IMG_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif',

Does it mean that I can use multi-band images (remote sensing)? For example, image with 4 channels (R,G,B,NIR), saved as image_0.tif.