Is the data out of bounds?

[opee007]

Hello, thanks for your nice work. The code ("depth = (depth*1000).astype(np.uint16)") is applied in the loaddata.py. As far as I know, the upper bound of uint16 is 65535, but some of the values of DSM multiplied by 1000 are greater than 65535 on the two ISPRS datasets, which I am confused about.

[speed8928]

Hi opee Have a peep in nyu_transform.py the line:294

depth = self.to_tensor(depth)/100000 When images converted to tensor I divided into 100000. If you want to try to train your own dataset please modify it to fit your purpose as not all DSM are in meter.

[opee007]

Hi, thank you a lot for your reply, the values of DSM on the Vaihingen dataset are about 200~300. For instance, the values output by the code below are incorrect. a = np.array(250*1000).astype(np.uint16) / 100000 print("a:", a)

a: 0.53392

In fact, a should be 2.5. Do you preprocess the original DSM, such as subtract a certain value from the original DSM on the Vaihingen dataset?

[speed8928]

Try a = np.array(250*1000).astype(np.uint16) a = a / 100000

yes, you need to normalize the DSM by subtracting DTM.

[Neroaway]

Do you directly use DSM Normalisation data in the ISPRS dataset for training? How do you convert this data unit into meters?