Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

support bg color #94

Draft
wants to merge 5 commits into
base: main
Choose a base branch
from
Draft

support bg color #94

wants to merge 5 commits into from

Conversation

wanmeihuali
Copy link
Owner

No description provided.

@github-actions
Copy link

Running experiment on sagemaker with git sha cd7417b

@github-actions
Copy link

@github-actions
Copy link

Running experiment on sagemaker with git sha cd7417b

@github-actions
Copy link

@github-actions
Copy link

Training job support-bg-color-cd7417b-230713-054228-tat-truck-baseline completed.

Model url: s3://taichi-3d-gaussian-splatting-log/tat-truck-baseline/support-bg-color-cd7417b-230713-054228-tat-truck-baseline/output/model.tar.gz,

tensorboard output path: s3://taichi-3d-gaussian-splatting-log/tat-truck-baseline/support-bg-color-cd7417b-230713-054228-tat-truck-baseline/output/output.tar.gz

@github-actions
Copy link

Training job support-bg-color-cd7417b-230713-054228-tat-truck-baseline final metrics:

Latest Metrics

train:iteration train:l1loss train:loss train:num_valid_points train:psnr train:ssim train:ssimloss val:loss val:psnr val:ssim
30000.0 0.028060728684067726 0.04665923863649368 463203.0 26.227439880371094 0.8789467215538025 0.12105327844619751 0.05423763766884804 24.982290267944336 0.864874541759491

Max Metrics

train:5kpsnr train:5kssim train:7kpsnr train:7kssim train:psnr train:ssim val:5kpsnr val:5kssim val:7kpsnr val:7kssim val:psnr val:ssim
22.77389907836914 0.811922550201416 25.415742874145508 0.8546379804611206 26.227439880371094 0.8789467215538025 22.970996856689453 0.8145796656608582 23.850791931152344 0.8371531963348389 24.982290267944336 0.864874541759491

@github-actions
Copy link

Training job support-bg-color-cd7417b-230713-054144-tat-train-baseline completed.

Model url: s3://taichi-3d-gaussian-splatting-log/tat-train-baseline/support-bg-color-cd7417b-230713-054144-tat-train-baseline/output/model.tar.gz,

tensorboard output path: s3://taichi-3d-gaussian-splatting-log/tat-train-baseline/support-bg-color-cd7417b-230713-054144-tat-train-baseline/output/output.tar.gz

@github-actions
Copy link

Training job support-bg-color-cd7417b-230713-054144-tat-train-baseline final metrics:

Latest Metrics

train:iteration train:l1loss train:loss train:num_valid_points train:psnr train:ssim train:ssimloss val:loss val:psnr val:ssim
30000.0 0.060022469609975815 0.08679809421300888 693872.0 21.207773208618164 0.8060994148254395 0.19390058517456055 0.10382827371358871 20.614830017089844 0.7850539088249207

Max Metrics

train:5kpsnr train:5kssim train:7kpsnr train:7kssim train:psnr train:ssim val:5kpsnr val:5kssim val:7kpsnr val:7kssim val:psnr val:ssim
22.61538314819336 0.7635285258293152 22.542423248291016 0.8190064430236816 21.207773208618164 0.8060994148254395 19.258577346801758 0.7346248030662537 19.708393096923828 0.7557579278945923 20.614830017089844 0.7850539088249207

@jb-ye
Copy link
Contributor

jb-ye commented Oct 25, 2023

Does this PR ready for merging (after rebasing)? Anything you plan to work on? @wanmeihuali

@wanmeihuali
Copy link
Owner Author

Does this PR ready for merging (after rebasing)? Anything you plan to work on? @wanmeihuali

The previous experiment does not show any significant improvement with background color...Actually, the PSNR even dropped a bit... I'm not sure if it will get better after rebasing. Anyway, we will only merge it if it does not hurt the metrics.

@jb-ye
Copy link
Contributor

jb-ye commented Oct 25, 2023

I prefer to use "black" background (0, 0, 0) as the default value for training. But at the render time, one can choose to switch between black or white or other custom colors.

@@ -1302,6 +1319,10 @@ def forward(self, input_data: GaussianPointCloudRasterisationInput):
q_pointcloud_camera = input_data.q_pointcloud_camera
t_pointcloud_camera = input_data.t_pointcloud_camera
color_max_sh_band = input_data.color_max_sh_band
background_color = input_data.background_color
if background_color is None:
background_color = torch.ones((3, ), dtype=torch.float32,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Set to zero by default

@github-actions
Copy link

Running experiment on sagemaker with git sha 3ddc747

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184713-garden created

@github-actions
Copy link

Running experiment on sagemaker with git sha 3ddc747

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184731-stump created

@github-actions
Copy link

Running experiment on sagemaker with git sha 3ddc747

@github-actions
Copy link

@github-actions
Copy link

Running experiment on sagemaker with git sha 3ddc747

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184805-bicycle created

@github-actions
Copy link

Running experiment on sagemaker with git sha 3ddc747

@github-actions
Copy link

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184731-stump completed.

Model url: s3://taichi-3d-gaussian-splatting-log/stump/support-bg-color-3ddc747-231025-184731-stump/output/model.tar.gz,

tensorboard output path: s3://taichi-3d-gaussian-splatting-log/stump/support-bg-color-3ddc747-231025-184731-stump/output/output.tar.gz

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184731-stump final metrics:

Latest Metrics

train:iteration train:l1loss train:loss train:psnr train:ssim train:ssimloss val:loss val:psnr val:ssim
30000.0 0.02718164585530758 0.06568731367588043 27.399700164794922 0.7802900075912476 0.21970999240875244 0.08477786183357239 25.340702056884766 0.7252534031867981

Max Metrics

train:5kpsnr train:5kssim train:7kpsnr train:7kssim train:psnr train:ssim val:5kpsnr val:5kssim val:7kpsnr val:7kssim val:psnr val:ssim
22.840965270996094 0.542497992515564 22.793359756469727 0.604921281337738 27.399700164794922 0.7802900075912476 23.599395751953125 0.6121237277984619 24.146839141845703 0.6494683027267456 25.340702056884766 0.7252534031867981

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184819-tat-truck-baseline completed.

Model url: s3://taichi-3d-gaussian-splatting-log/tat-truck-baseline/support-bg-color-3ddc747-231025-184819-tat-truck-baseline/output/model.tar.gz,

tensorboard output path: s3://taichi-3d-gaussian-splatting-log/tat-truck-baseline/support-bg-color-3ddc747-231025-184819-tat-truck-baseline/output/output.tar.gz

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184819-tat-truck-baseline final metrics:

Latest Metrics

train:iteration train:l1loss train:loss train:psnr train:ssim train:ssimloss val:loss val:psnr val:ssim
30000.0 0.02906094305217266 0.048246562480926514 26.5162296295166 0.8750109672546387 0.12498903274536133 0.05499852076172829 25.008712768554688 0.8628195524215698

Max Metrics

train:5kpsnr train:5kssim train:7kpsnr train:7kssim train:psnr train:ssim val:5kpsnr val:5kssim val:7kpsnr val:7kssim val:psnr val:ssim
23.006772994995117 0.8178406953811646 24.454757690429688 0.8571473360061646 26.5162296295166 0.8750109672546387 23.024946212768555 0.8145551681518555 23.754863739013672 0.8370049595832825 25.008712768554688 0.8628195524215698

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184805-bicycle completed.

Model url: s3://taichi-3d-gaussian-splatting-log/bicycle/support-bg-color-3ddc747-231025-184805-bicycle/output/model.tar.gz,

tensorboard output path: s3://taichi-3d-gaussian-splatting-log/bicycle/support-bg-color-3ddc747-231025-184805-bicycle/output/output.tar.gz

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184805-bicycle final metrics:

Latest Metrics

train:iteration train:l1loss train:loss train:psnr train:ssim train:ssimloss val:loss val:psnr val:ssim
30000.0 0.043804850429296494 0.09135077893733978 21.698631286621094 0.7184655070304871 0.28153449296951294 0.09074745327234268 24.536022186279297 0.7150701880455017

Max Metrics

train:5kpsnr train:5kssim train:7kpsnr train:7kssim train:psnr train:ssim val:5kpsnr val:5kssim val:7kpsnr val:7kssim val:psnr val:ssim
24.073598861694336 0.6494255065917969 23.9046688079834 0.6313458681106567 21.698631286621094 0.7184655070304871 22.656890869140625 0.5860998630523682 23.194732666015625 0.6250209212303162 24.536022186279297 0.7150701880455017

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184754-tat-train-baseline completed.

Model url: s3://taichi-3d-gaussian-splatting-log/tat-train-baseline/support-bg-color-3ddc747-231025-184754-tat-train-baseline/output/model.tar.gz,

tensorboard output path: s3://taichi-3d-gaussian-splatting-log/tat-train-baseline/support-bg-color-3ddc747-231025-184754-tat-train-baseline/output/output.tar.gz

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184754-tat-train-baseline final metrics:

Latest Metrics

train:iteration train:l1loss train:loss train:psnr train:ssim train:ssimloss val:loss val:psnr val:ssim
30000.0 0.057057030498981476 0.07781828194856644 21.757766723632812 0.8391367197036743 0.16086328029632568 0.10022936016321182 20.933279037475586 0.8011162877082825

Max Metrics

train:5kpsnr train:5kssim train:7kpsnr train:7kssim train:psnr train:ssim val:5kpsnr val:5kssim val:7kpsnr val:7kssim val:psnr val:ssim
22.54574203491211 0.8242156505584717 24.459022521972656 0.8401768207550049 21.757766723632812 0.8391367197036743 19.247770309448242 0.7367949485778809 19.85987091064453 0.7608734965324402 20.933279037475586 0.8011162877082825

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184713-garden completed.

Model url: s3://taichi-3d-gaussian-splatting-log/garden/support-bg-color-3ddc747-231025-184713-garden/output/model.tar.gz,

tensorboard output path: s3://taichi-3d-gaussian-splatting-log/garden/support-bg-color-3ddc747-231025-184713-garden/output/output.tar.gz

@github-actions
Copy link

Training job support-bg-color-3ddc747-231025-184713-garden final metrics:

Latest Metrics

train:iteration train:l1loss train:loss train:psnr train:ssim train:ssimloss val:loss val:psnr val:ssim
30000.0 0.042713962495326996 0.08603265881538391 22.841739654541016 0.7406925559043884 0.2593074440956116 0.06810751557350159 26.158889770507812 0.7931697964668274

Max Metrics

train:5kpsnr train:5kssim train:7kpsnr train:7kssim train:psnr train:ssim val:5kpsnr val:5kssim val:7kpsnr val:7kssim val:psnr val:ssim
26.06695556640625 0.7288302183151245 26.439626693725586 0.7657265067100525 22.841739654541016 0.7406925559043884 24.464326858520508 0.708663284778595 25.03775978088379 0.7400090098381042 26.158889770507812 0.7931697964668274

@@ -652,6 +660,8 @@ def gaussian_point_rasterisation_backward(
# \frac{dC}{da_i} = c_i T(i) - \frac{1}{1 - a_i} w_i
alpha_grad_from_rgb = (color * T_i - w_i / (1. - alpha)) \
* pixel_rgb_grad
alpha_grad_from_rgb -= pixel_rgb_grad * background_color_vector * \
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you provide the formula to calculate the gradient properly? I am a bit confused by the name convention.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants