likyoo / multimodal-remote-sensing-toolkit Goto Github PK

Dear @likyoo,

I have downloaded the Houston 2013 dataset from the official page.

Equipped with osgeo.gdal and osgeo.gdal_array.DatasetReadAsArray packages, I got HSI.mat and LiDAR.mat just as what's inside the google drive url, just as follows:

from osgeo import gdal
from osgeo.gdal_array import DatasetReadAsArray
HSI = gdal.Open('2013_IEEE_GRSS_DF_Contest_CASI.tif')
HSI = DatasetReadAsArray(HSI)
io.savemat('HSI.mat', {'HSI': HSI})

To my disappointment, these files don't share the same contents.

Could you please tell me how you did the preprocessing step?

Best regards

Hi, likyoo!

Thank you so much for such a solid code! I appreciate the repository very much.

When I am reading the function mixture_noise within datasets.py as follows:

    def mixture_noise(self, data, label, beta=1 / 25):
        alpha1, alpha2 = np.random.uniform(0.01, 1.0, size=2)
        noise = np.random.normal(loc=0.0, scale=1.0, size=data.shape)
        data2 = np.zeros_like(data)
        for idx, value in np.ndenumerate(label):
            if value not in self.ignored_labels:
                l_indices = np.nonzero(self.labels == value)[0]
                l_indice = np.random.choice(l_indices)
                assert self.labels[l_indice] == value
                x, y = self.indices[l_indice]
                data2[idx] = self.data[x, y]
        return (alpha1 * data + alpha2 * data2) / (alpha1 + alpha2) + beta * noise

I found the following question:

• self.labels is of the type list.
• l_indices = np.nonzero(self.labels == value)[0] may be not in accordance with the intent.

However, although I fix if self.mixture_augmentation and np.random.random() < 0.2 to if self.mixture_augmentation and run the script with --mixture-augmentation, the training procedure runs without exception, which confused me a lot.

Another question I enountered is:
I notice the flip_augmentation apply on LiDAR and HSI data both. But The radiation_augmentation and mixture_augmentation only apply on the HSI data, leaving LiDAR data alone.
Could you please tell me the reason behind?

Best wishes!

If no padding is applied on img1, img2, and gt, many pixels will be classified as Unlabeled in the test stage.

Besides, it will also cause the classification map to lose the border, with a width of patch_size//2

can you please delete this issue. I created it by mistake. sorry