RuntimeError about deepdepthdenoising

[spthermo]

Hi,

the model supports 640x360 input resolution as input. Since you use a different resolution, I presume that the error is at the concatenation of an interpolation layer output with the features of the corresponding skip connection. This happens because the interpolation layers utilize the default h and w (360 and 640, see models/shallow_partial.py), while the skip connections feed the decoder with feature maps of your h' and w', where h' < h and w' < w (i.e. 15 < 90 in your case).

[Harleens13]

Hi @spthermo. Thank for the reply.
I get this part. But sometimes an image of 1280*720 has also been read but an image with dimensions 694*512 gives errors.
Also I have found another discrepancy, when my input image is in .jpg format, I am able to see the denoised depth image (output from inference.py) but can't see the denoised depth image when it is in .png format (it is totally black image everytime). Interestingly the pointclouds in both the cases are available and somewhat similar. Did you guys had the same observation while testing?

[kearcer]

@spthermo yes,i also found input the png image will output the all black images,how to solve this?

[zapaishchykova]

@kearcer try to download another image viewer, like ImageJ

[zuru]

After loading a file with the load_depth function, the result should be an image tensor containing the depth values in meters (the scale parameter is responsible for that scaling, with the default value being 0.001 and assuming that the image contains the depth in millimeters).

Typical formats for depth images would be:

• .exr with raw float values (in arbitrary scaling, in the case of meters for example, it should be load_depth(filename, scale=1.0))
• .png with unsigned short (i.e. 16bit) values (usually in millimeters)
• .pgm with unsigned short (i.e. 16bit) values (usually in millimeters)

It should be noted that .jpeg files are unsuitable to store depth values.

If you need more help with your files, please upload a sample that we can check.