bingyaohuang / compennet Goto Github PK

I really admire the author's work ！

After reading your article, I saw that you used TPS
as a comparison method. Can you open source the
code used by TPS on your dataset?

And I saw from Issues that you mentioned that TPS uses all
The shading map of, which makes me feel a little confused,
your dataset only has the last shading capture image.

Looking forward to your reply, I will be grateful .

The execution bails out at the end of the first iteration with the above error:

How to fix it:
everywhere you call .to(device) you need to change it to = .to(device).

The "to" function call isn't modifying the object it's called upon.

did you capture RAW images and then use a photoediting software to generate gamma corrected png images?

The compensated images obtained after inferencing lose their sharpness. As I understand, autoencoders due to their nature of encoding and then decoding, will result in the loss of detail. Do you observe this behavior as well in your experiments? Is it possible to somehow reduce this loss? A way that we preserve the edge detail or perhaps combine the original and the generated compensation image to get the best of both worlds.

I am curious as to what the results in your lab looked like with larger resolutions. Although compensation works to hide the screen imperfections, the loss in sharpness results in an objective observer preferring the sharp uncompensated image over a compensated image.

I ran your code with my setup and attached is the result for one of the test images.

I see that the images expected, post compensation look very good and the color patches I added almost completely disappear. However, when I project the compensation images from the prj folder and capture the images, the results don't look anywhere close to the expectation.

I don't think it's because of the solid color patches being too saturated because in your testing, you used screens that were just as saturated.

I am using a pretty old projector (Powerlite 7900P) projecting at 800x600 and an Intel realsense RGB camera capturing at 1920x1080. Fixed exposure, fixed Whitebalance.
I trained with 100 images, l1+ssim, 1000 iterations

img_gray_color - solid gray projected on my screen
img_0006 - original image
img_0006_Color - capture of the original image projected on screen
img_0006_Color_corrected - capture of the compensation image on screen
img_0006_generated - generated by the model as the expected output post compensation

How did the results of projecting the compensated image look like in your lab? Were they as good as the generated output expected it to be?
Would using a more modern projector give better results when projecting a compensated image? (I assume it should. Given that the contrast ratio would be vastly better)

File "main.py", line 172, in
compen_net.module.name = model_name
File "/home/spoluri/cm/CompenNet/virtual_env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 539, in getattr
type(self).name, name))
AttributeError: 'CompenNet' object has no attribute 'module'

Solution:
Remove module from line 172 and it should look like this: compenNet.name = model_name

Hello ,I'm training the original dataset, and I want to shot some pics for calibration, so I want to know whether the dataset is real shot or sythesized, and should the image be adjusted to look like the one in the data set?