gabrieleilertsen / hdrcnn

HDR image reconstruction from a single exposure using deep CNNs
https://computergraphics.on.liu.se/hdrcnn/
BSD 3-Clause "New" or "Revised" License
502 stars 101 forks source link

Does the images in the folder “data” need to be pre-processed? #12

Closed cherryMo closed 5 years ago

cherryMo commented 6 years ago

Does the images in the folder data need to be preprocessed?

Why is it not good to process images downloaded internet? the image from internet img04 _in.png 000001_in _out.png 000001_out

gabrieleilertsen commented 6 years ago

Hi. No pre-processing required. I would guess that the problem is that you used the weights trained without compression included, is that the case? The image probably experiences compression artifacts, which requires the weights trained with such artifacts included (the correct weights can be downloaded here).

cherryMo commented 5 years ago

Thank you for your reply, but the problem has not been resolved. Processing my own image, the effect is not very good (the use of weights is correct)

gabrieleilertsen commented 5 years ago

Are you sure about the weights? I tried using your image above, and get the exact same result when using the weights that do not include compression (hdrcnn_params.npz). Using the weights that are trained with compression (hdrcnn_params_compr.npz), the result looks like this:

000001_out

The command for running the reconstruction:

$ python hdrcnn_predict.py --params hdrcnn_params_compr.npz --im_dir input.jpg --width 1280 --height 852
cherryMo commented 5 years ago

Thank you for your reply. could you explain the difference between input_image, _in.png _out.png and out_hdr in detail?

gabrieleilertsen commented 5 years ago

The input image is your input low dynamic range image for which you want to reconstruct HDR information. For the output images, please see the explanation in a previous issue comment.