Closed haoz19 closed 1 month ago
Hi! To calculate metrics for the test, please use calculate_metrics.py
Thanks. I noticed a problem with the male-3-casual training config. Committed the corrected version
Thanks for the instruction!
Should we use this:
python calculate_metrics.py --ground_truth <gt_folder> --predict <pred_folder>
gt folder: rgb_image in test folder pred_folder: rasterization in test folder?
Thanks!
@haoz19 Did you get the evaluation result same with paper?
When I ran the calculate_metrics the result is really low.
python calculate_metrics.py --predict ./logs/gaussians_docker_female3/female-3-casual-2024_06-05_16-10-test/test/rasterization --ground_truth ./logs/gaussians_docker_female3/female-3-casual-2024_06-05_16-10-test/test/rgb_image
male3:
psnr: tensor(14.6720, device='cuda:0')
ssim: tensor(0.8615, device='cuda:0')
lpips: tensor(0.1739, device='cuda:0')
male4:
psnr: tensor(15.7759, device='cuda:0')
ssim: tensor(0.8782, device='cuda:0')
lpips: tensor(0.1730, device='cuda:0')
female3:
psnr: tensor(13.6854, device='cuda:0')
ssim: tensor(0.8767, device='cuda:0')
lpips: tensor(0.1674, device='cuda:0')
Okay, let's localize the problem. Check that the background on GT and generated images is the same color (black). Then, ensure that the metrics from the pre-trained downloaded checkpoint match the article's ones. To understand if the problem in measuring metrics or training
Hi @duy-maimanh ,
I got results like this:
00016: psnr: tensor(27.3044, device='cuda:0') ssim: tensor(0.9398, device='cuda:0') lpips: tensor(0.0399, device='cuda:0')
male-3: psnr: tensor(26.5395, device='cuda:0') ssim: tensor(0.9617, device='cuda:0') lpips: tensor(0.0376, device='cuda:0')
male-4: psnr: tensor(24.6781, device='cuda:0') ssim: tensor(0.9438, device='cuda:0') lpips: tensor(0.0567, device='cuda:0')
@david-svitov @haoz19 Thank you very much for your answer.
Hi,
Thanks for the great work!
I ran the training code, and also the rendering code for male3-casual, and I saw that it shows psnr : 24.94 during training, and during testing, it doesn't show psnr. While it is lower than reported in the paper which is 31.46.
I would like to ask if the issue is caused by the parameters I used for training or some other reasons.
Many Thanks, Hao