Some problem with PSNR calculating

[tingzhongyue]

Thank you for your great work and contribution.

When I tried to train 4dgs, I found the results saved by .pth are inconsistent with the evaluation results during training. For example, in sear_steak, the PSNR of 4dgs at epoch 6000 when training is 33.02, but when I load pth for testing, it gives me 32.63.

My test code is written according to this issuehttps://github.com/fudan-zvg/4d-gaussian-splatting/issues/12.

And PSNR evaluation code is shown below

psnr_test = 0.0
for idx, batch_data in enumerate(tqdm(views, desc="Rendering progress")):
    gt_image, viewpoint = batch_data
    gt_image = gt_image.cuda()
    viewpoint = viewpoint.cuda()
    render_pkg = render(viewpoint, gaussians, pipeline, background)
    image = torch.clamp(render_pkg["render"], 0.0, 1.0)
    psnr_test += psnr(image, gt_image).mean().double()
psnr_test /= len(views)
print(psnr_test)

Do you ever meet some problems like this, and do you know any method to solve it?

[tingzhongyue]

It is worth mentioning that when I tested my code on dnerf, it worked well. And on the dynerf dataset, it also worked when it was 3000th iteration. However, once the iterations increased, this problem would appear. And this also happens when I try to test the saved ply file

[Alexander0Yang]

Hi, @tingzhongyue,

I haven't encountered this issue. For coffee_martini and flame_salmon, the testing script provided in #12 might cause a significant drop in performance because it does not load the envmap, but for other scenes, this script is sufficient to reproduce the performance reported in the paper, as mentioned in #24.

Actually, you can check the rendered images and the input parameters of the rasterizer in both the training process and your own testing scripts to provide more information for identifying the problem.