does the imgs with _t.png is the final reslut ?

[zhangyunming]

as said: The test results will be saved to ./results/testset_name.

does the imgs with _t.png is the final reslut ?

[chxy95]

as said: The test results will be saved to ./results/testset_name.

does the imgs with _t.png is the final reslut ?

I cannot understand your question, but the test result of an image should be a reconstructed image with a align ratio encoded in a .npy file. The naming rules of this repo comply with the requirements of the NTIRE competition: example: for an input single image with filename ["0003_medium.png"] the output file should be ["0003.png", "0003_alignratio.npy"]. If you have any question about the naming relus or you want to change the rules, you can check the file ./codes/test.py of line 57-58 and modify the related functions in ./codes/utils/util.py.

[chxy95]

@zhangyunming
It looks weird...
The HDR results need to be processed by tone-mapping for visualization.
Nonetheless, you'd better use the dataset provided by the competition organizer for validation.
Otherwise, you may get strange results due to the large difference in data distribution.

[zhangyunming]

indeed , it looks weird,

the hdr result is get by :

any problem?

it is myown data.

[chxy95]

@zhangyunming The code of visualization is for the competition data. It requires the Ground Truth to calculate parameters of the tone-mapping algorithm. You may use the tanh function to make an approximate visualization for your own data.

img = io.imread_uint16_png(img_path, alignratio_path)
img = np.clip(np.tanh(img), 0, 1) * 255
img = img.round().astype(np.uint8)
cv2.imwrite(save_path, cv2.cvtColor(img, cv2.COLOR_RGB2BGR))

[zhangyunming]

use your code and the hdr result is all balck

[zhangyunming]

for own data:

[chxy95]

use your code and the hdr result is all balck

I edited the code and you can check it again. (forgot to multiply by 255.)

[zhangyunming]

actually , i have *255, but the result still black , the pix is ~3-5, so is black.
i use the code ：

hdr_image = io.imread_uint16_png('F:\zym_AI_tests\HDRUNet\codes\imgs\(121)ev-.jpg', 'F:\zym_AI_tests\HDRUNet\results\1\1\(121_alignratio.npy')
hdr_linear_image = hdr_image ** 3.8
norm_perc = np.percentile(hdr_linear_image, 99)

hdr_image = (m.tanh_norm_mu_tonemap(hdr_linear_image, norm_perc) * 255.).round().astype(np.uint8)
cv2.imshow("hdr_demo", cv2.cvtColor(hdr_image, cv2.COLOR_RGB2BGR))
cv2.imwrite('F:\zym_AI_tests\HDRUNet\codes\imgs\(121)ev-hdr.png', cv2.cvtColor(hdr_image, cv2.COLOR_RGB2BGR))
cv2.waitKey(0)

and the result above seems correct.

[chxy95]

2.24 in hdr_linear_image = hdr_image ** 2.24 is the Gamma value, so it is unreasonable to change it into 3.8.
norm_perc is the key parameters calculated using the Ground Truth, and it's fine to change this value for a good view.

[zhangyunming]

strange, i use your code:

img = io.imread_uint16_png('F:\zym_AI_tests\HDRUNet\codes\imgs\(121)ev-.jpg', 'F:\zym_AI_tests\HDRUNet\results\1\1\(121_alignratio.npy')
img = np.clip(np.tanh(img), 0, 1) * 255
img = img.round().astype(np.uint8)
cv2.imshow("hdr", cv2.cvtColor(img, cv2.COLOR_RGB2BGR))
print(img)
#cv2.imwrite(save_path, cv2.cvtColor(img, cv2.COLOR_RGB2BGR))
cv2.waitKey(0)

still black! print(img)
[0 0 1]
[0 0 0]
...
[3 3 3]
[3 3 3]
[3 3 3]]

...

[[1 1 1]
[1 1 1]
[1 1 1]
...
[1 1 1]
[1 1 1]
[1 1 1]]

[[1 1 1]
[1 1 1]
[1 1 1]
...
[1 1 1]
[1 1 1]
[1 1 1]]

[[1 1 1]
[1 1 1]
[1 1 1]
...
[1 1 1]
[1 1 1]
[1 1 1]]]

[chxy95]

You can send the input image and the corresponding reconstructed image with the align ratio to this e-mail [email protected]. I will check it in my free time.