pretrained model seems get poor PSNR and mu-PSNR about twostagehdr_ntire21 HOT 4 CLOSED

Can you please tell me which exposure setting you are using? According to the challenge protocol, the model was trained with medium exposure images only.

from twostagehdr_ntire21.

I've cross-checked the output with the pretrained weights. It seems to be fine. For your reference, I've added sample outputs in the modelOutput/ directory. . Please check your input images.

from twostagehdr_ntire21.

I've cross-checked the output with the pretrained weights. It seems to be fine. For your reference, I've added sample outputs in the modelOutput/ directory. . Please check your input images.

Hi !
I use the images in modelOutput/Two-stageHDR/HDRInput/ and the provided pretrained model to generate output, and I got
0028_alignratio.npy

np.load('./testOutTwo-stageHDR/0028/0028_alignratio.npy').astype(np.float32)
array(19877.906, dtype=float32)
while in your provided output,
np.load('./modelOutput/Two-stageHDR/HDRInput/0028_alignratio.npy').astype(np.float32)
array(21488.842, dtype=float32)

I didn't change any code

from twostagehdr_ntire21.

Hello!
xxx_alignration.py is metadata for normalizing 16-bit images. Notably, the input of our model is 8-bit images and any 16bit-samples/.npy files have to be ignored while reading data input images.

The samples we have provided in the "modelOutput/" directory are saved according to the NTIRE-21 submission protocol. Ideally, these samples should not be used for the inference. I will strongly suggest you to go through the following link: https://competitions.codalab.org/competitions/28161. You may also download the testing samples from the given link as well (after registration). It is worth noting, the output of our Model is stored in 16-bit format, which is the range between (0-65536). Therefore, we may not treat them as normal images. Please refer to the original paper and the competition link for the details. Also, for calculating PSNR and mu-PSNR we should use the script provided by the organizer of the competition. However, for any difficulties feel free to contact us.

from twostagehdr_ntire21.